Release whisper-base-mla-cv11: on-device tier of the MLA-Whisper family

README.md

@@ -0,0 +1,58 @@
+---
+license: apache-2.0
+base_model: openai/whisper-base
+datasets:
+- burakaydinofficial/Whispered
+language: [en, de, es, fr, ru, tr, cy, ar, ka, th, zh]
+pipeline_tag: automatic-speech-recognition
+library_name: transformers
+tags: [whisper, mla, multi-head-latent-attention, kv-cache-compression, speech-recognition, multilingual]
+---
+
+# Whisper-Base-MLA (11 languages) — MLA-converted, 62.5% smaller decode KV-cache
+
+The on-device-tier sibling of the WhisperMLA family: `openai/whisper-base` (74M) with its
+decoder self-attention converted MHA→MLA (per [Whisper-MLA, arXiv:2603.00563](https://arxiv.org/abs/2603.00563)),
+recovery-fine-tuned on 11 languages of the CC0
+[Whispered](https://huggingface.co/datasets/burakaydinofficial/Whispered) corpus (32k clips/lang).
+
+```python
+from transformers import AutoModelForSpeechSeq2Seq
+model = AutoModelForSpeechSeq2Seq.from_pretrained("burakaydinofficial/whisper-base-mla-cv11", trust_remote_code=True)  # transformers==4.46.x
+```
+
+## Honest sizing note (read this first)
+**Conversion cost grows as the student shrinks** — measured across the family: small ≈ +0.4
+median WER → base ≈ +1.0 → tiny ≈ +1.9. At the base tier you pay ≈ +1.0 WER (median) for the
+62.5% cache cut. If quality is the priority, prefer the `small` variant; this tier is for
+memory-constrained deployments where the cache cut matters most.
+
+## Results (CommonVoice-17 test, n=1500/lang; WER/CER %; cost = paired vs an identically-trained unconverted control)
+
+| Lang | this model (WER / CER) | conversion cost |
+|---|---|---|
+| en | 21.1 / 11.2 | +1.04 ✱ |
+| de | 30.1 / 11.3 | +0.83 (ns) |
+| es | 19.0 / 6.7 | +0.88 ✱ |
+| fr | 32.3 / 13.9 | +0.59 (ns) |
+| ru | 28.8 / 8.7 | +1.31 ✱ |
+| tr | 38.9 / 11.8 | +1.13 ✱ |
+| cy | 79.9 / 34.2 | +2.10 ✱ |
+| ar | 58.9 / 24.5 | +2.71 ✱ |
+| th | 51.7 / 20.3 | +0.22 CER (ns) |
+| zh | 91.7 / 25.4 | −0.23 CER (ns) |
+| ka | 111.8 / 75.8 | +12.74 ✱ — floor |
+
+Absolute quality is base-tier-typical (much lower than `small` — that is the base model, not
+MLA). Encoder frozen both arms; 15,000 steps; warmup+cosine; fp16.
+
+## Limitations
+Same as the flagship: transformers==4.46.x + trust_remote_code required; not loadable in
+whisper.cpp/faster-whisper/CT2; coverage = these 11 languages (unseen scripts degrade);
+ka reported as the labeled model-class floor; read-speech domain.
+
+## Acoustic conditions of the evaluation
+Evaluated on crowdsourced **consumer-microphone recordings with real environmental noise** —
+traffic, room reverb, variable devices — CommonVoice's native conditions, not studio audio.
+The numbers above already include that heterogeneity. Not yet benchmarked: far-field, telephony
+(8 kHz), overlapping speech; an SNR-ladder robustness section will be added when measured.

added_tokens.json

@@ -0,0 +1,1609 @@
+{
+  "<|0.00|>": 50364,
+  "<|0.02|>": 50365,
+  "<|0.04|>": 50366,
+  "<|0.06|>": 50367,
+  "<|0.08|>": 50368,
+  "<|0.10|>": 50369,
+  "<|0.12|>": 50370,
+  "<|0.14|>": 50371,
+  "<|0.16|>": 50372,
+  "<|0.18|>": 50373,
+  "<|0.20|>": 50374,
+  "<|0.22|>": 50375,
+  "<|0.24|>": 50376,
+  "<|0.26|>": 50377,
+  "<|0.28|>": 50378,
+  "<|0.30|>": 50379,
+  "<|0.32|>": 50380,
+  "<|0.34|>": 50381,
+  "<|0.36|>": 50382,
+  "<|0.38|>": 50383,
+  "<|0.40|>": 50384,
+  "<|0.42|>": 50385,
+  "<|0.44|>": 50386,
+  "<|0.46|>": 50387,
+  "<|0.48|>": 50388,
+  "<|0.50|>": 50389,
+  "<|0.52|>": 50390,
+  "<|0.54|>": 50391,
+  "<|0.56|>": 50392,
+  "<|0.58|>": 50393,
+  "<|0.60|>": 50394,
+  "<|0.62|>": 50395,
+  "<|0.64|>": 50396,
+  "<|0.66|>": 50397,
+  "<|0.68|>": 50398,
+  "<|0.70|>": 50399,
+  "<|0.72|>": 50400,
+  "<|0.74|>": 50401,
+  "<|0.76|>": 50402,
+  "<|0.78|>": 50403,
+  "<|0.80|>": 50404,
+  "<|0.82|>": 50405,
+  "<|0.84|>": 50406,
+  "<|0.86|>": 50407,
+  "<|0.88|>": 50408,
+  "<|0.90|>": 50409,
+  "<|0.92|>": 50410,
+  "<|0.94|>": 50411,
+  "<|0.96|>": 50412,
+  "<|0.98|>": 50413,
+  "<|1.00|>": 50414,
+  "<|1.02|>": 50415,
+  "<|1.04|>": 50416,
+  "<|1.06|>": 50417,
+  "<|1.08|>": 50418,
+  "<|1.10|>": 50419,
+  "<|1.12|>": 50420,
+  "<|1.14|>": 50421,
+  "<|1.16|>": 50422,
+  "<|1.18|>": 50423,
+  "<|1.20|>": 50424,
+  "<|1.22|>": 50425,
+  "<|1.24|>": 50426,
+  "<|1.26|>": 50427,
+  "<|1.28|>": 50428,
+  "<|1.30|>": 50429,
+  "<|1.32|>": 50430,
+  "<|1.34|>": 50431,
+  "<|1.36|>": 50432,
+  "<|1.38|>": 50433,
+  "<|1.40|>": 50434,
+  "<|1.42|>": 50435,
+  "<|1.44|>": 50436,
+  "<|1.46|>": 50437,
+  "<|1.48|>": 50438,
+  "<|1.50|>": 50439,
+  "<|1.52|>": 50440,
+  "<|1.54|>": 50441,
+  "<|1.56|>": 50442,
+  "<|1.58|>": 50443,
+  "<|1.60|>": 50444,
+  "<|1.62|>": 50445,
+  "<|1.64|>": 50446,
+  "<|1.66|>": 50447,
+  "<|1.68|>": 50448,
+  "<|1.70|>": 50449,
+  "<|1.72|>": 50450,
+  "<|1.74|>": 50451,
+  "<|1.76|>": 50452,
+  "<|1.78|>": 50453,
+  "<|1.80|>": 50454,
+  "<|1.82|>": 50455,
+  "<|1.84|>": 50456,
+  "<|1.86|>": 50457,
+  "<|1.88|>": 50458,
+  "<|1.90|>": 50459,
+  "<|1.92|>": 50460,
+  "<|1.94|>": 50461,
+  "<|1.96|>": 50462,
+  "<|1.98|>": 50463,
+  "<|10.00|>": 50864,
+  "<|10.02|>": 50865,
+  "<|10.04|>": 50866,
+  "<|10.06|>": 50867,
+  "<|10.08|>": 50868,
+  "<|10.10|>": 50869,
+  "<|10.12|>": 50870,
+  "<|10.14|>": 50871,
+  "<|10.16|>": 50872,
+  "<|10.18|>": 50873,
+  "<|10.20|>": 50874,
+  "<|10.22|>": 50875,
+  "<|10.24|>": 50876,
+  "<|10.26|>": 50877,
+  "<|10.28|>": 50878,
+  "<|10.30|>": 50879,
+  "<|10.32|>": 50880,
+  "<|10.34|>": 50881,
+  "<|10.36|>": 50882,
+  "<|10.38|>": 50883,
+  "<|10.40|>": 50884,
+  "<|10.42|>": 50885,
+  "<|10.44|>": 50886,
+  "<|10.46|>": 50887,
+  "<|10.48|>": 50888,
+  "<|10.50|>": 50889,
+  "<|10.52|>": 50890,
+  "<|10.54|>": 50891,
+  "<|10.56|>": 50892,
+  "<|10.58|>": 50893,
+  "<|10.60|>": 50894,
+  "<|10.62|>": 50895,
+  "<|10.64|>": 50896,
+  "<|10.66|>": 50897,
+  "<|10.68|>": 50898,
+  "<|10.70|>": 50899,
+  "<|10.72|>": 50900,
+  "<|10.74|>": 50901,
+  "<|10.76|>": 50902,
+  "<|10.78|>": 50903,
+  "<|10.80|>": 50904,
+  "<|10.82|>": 50905,
+  "<|10.84|>": 50906,
+  "<|10.86|>": 50907,
+  "<|10.88|>": 50908,
+  "<|10.90|>": 50909,
+  "<|10.92|>": 50910,
+  "<|10.94|>": 50911,
+  "<|10.96|>": 50912,
+  "<|10.98|>": 50913,
+  "<|11.00|>": 50914,
+  "<|11.02|>": 50915,
+  "<|11.04|>": 50916,
+  "<|11.06|>": 50917,
+  "<|11.08|>": 50918,
+  "<|11.10|>": 50919,
+  "<|11.12|>": 50920,
+  "<|11.14|>": 50921,
+  "<|11.16|>": 50922,
+  "<|11.18|>": 50923,
+  "<|11.20|>": 50924,
+  "<|11.22|>": 50925,
+  "<|11.24|>": 50926,
+  "<|11.26|>": 50927,
+  "<|11.28|>": 50928,
+  "<|11.30|>": 50929,
+  "<|11.32|>": 50930,
+  "<|11.34|>": 50931,
+  "<|11.36|>": 50932,
+  "<|11.38|>": 50933,
+  "<|11.40|>": 50934,
+  "<|11.42|>": 50935,
+  "<|11.44|>": 50936,
+  "<|11.46|>": 50937,
+  "<|11.48|>": 50938,
+  "<|11.50|>": 50939,
+  "<|11.52|>": 50940,
+  "<|11.54|>": 50941,
+  "<|11.56|>": 50942,
+  "<|11.58|>": 50943,
+  "<|11.60|>": 50944,
+  "<|11.62|>": 50945,
+  "<|11.64|>": 50946,
+  "<|11.66|>": 50947,
+  "<|11.68|>": 50948,
+  "<|11.70|>": 50949,
+  "<|11.72|>": 50950,
+  "<|11.74|>": 50951,
+  "<|11.76|>": 50952,
+  "<|11.78|>": 50953,
+  "<|11.80|>": 50954,
+  "<|11.82|>": 50955,
+  "<|11.84|>": 50956,
+  "<|11.86|>": 50957,
+  "<|11.88|>": 50958,
+  "<|11.90|>": 50959,
+  "<|11.92|>": 50960,
+  "<|11.94|>": 50961,
+  "<|11.96|>": 50962,
+  "<|11.98|>": 50963,
+  "<|12.00|>": 50964,
+  "<|12.02|>": 50965,
+  "<|12.04|>": 50966,
+  "<|12.06|>": 50967,
+  "<|12.08|>": 50968,
+  "<|12.10|>": 50969,
+  "<|12.12|>": 50970,
+  "<|12.14|>": 50971,
+  "<|12.16|>": 50972,
+  "<|12.18|>": 50973,
+  "<|12.20|>": 50974,
+  "<|12.22|>": 50975,
+  "<|12.24|>": 50976,
+  "<|12.26|>": 50977,
+  "<|12.28|>": 50978,
+  "<|12.30|>": 50979,
+  "<|12.32|>": 50980,
+  "<|12.34|>": 50981,
+  "<|12.36|>": 50982,
+  "<|12.38|>": 50983,
+  "<|12.40|>": 50984,
+  "<|12.42|>": 50985,
+  "<|12.44|>": 50986,
+  "<|12.46|>": 50987,
+  "<|12.48|>": 50988,
+  "<|12.50|>": 50989,
+  "<|12.52|>": 50990,
+  "<|12.54|>": 50991,
+  "<|12.56|>": 50992,
+  "<|12.58|>": 50993,
+  "<|12.60|>": 50994,
+  "<|12.62|>": 50995,
+  "<|12.64|>": 50996,
+  "<|12.66|>": 50997,
+  "<|12.68|>": 50998,
+  "<|12.70|>": 50999,
+  "<|12.72|>": 51000,
+  "<|12.74|>": 51001,
+  "<|12.76|>": 51002,
+  "<|12.78|>": 51003,
+  "<|12.80|>": 51004,
+  "<|12.82|>": 51005,
+  "<|12.84|>": 51006,
+  "<|12.86|>": 51007,
+  "<|12.88|>": 51008,
+  "<|12.90|>": 51009,
+  "<|12.92|>": 51010,
+  "<|12.94|>": 51011,
+  "<|12.96|>": 51012,
+  "<|12.98|>": 51013,
+  "<|13.00|>": 51014,
+  "<|13.02|>": 51015,
+  "<|13.04|>": 51016,
+  "<|13.06|>": 51017,
+  "<|13.08|>": 51018,
+  "<|13.10|>": 51019,
+  "<|13.12|>": 51020,
+  "<|13.14|>": 51021,
+  "<|13.16|>": 51022,
+  "<|13.18|>": 51023,
+  "<|13.20|>": 51024,
+  "<|13.22|>": 51025,
+  "<|13.24|>": 51026,
+  "<|13.26|>": 51027,
+  "<|13.28|>": 51028,
+  "<|13.30|>": 51029,
+  "<|13.32|>": 51030,
+  "<|13.34|>": 51031,
+  "<|13.36|>": 51032,
+  "<|13.38|>": 51033,
+  "<|13.40|>": 51034,
+  "<|13.42|>": 51035,
+  "<|13.44|>": 51036,
+  "<|13.46|>": 51037,
+  "<|13.48|>": 51038,
+  "<|13.50|>": 51039,
+  "<|13.52|>": 51040,
+  "<|13.54|>": 51041,
+  "<|13.56|>": 51042,
+  "<|13.58|>": 51043,
+  "<|13.60|>": 51044,
+  "<|13.62|>": 51045,
+  "<|13.64|>": 51046,
+  "<|13.66|>": 51047,
+  "<|13.68|>": 51048,
+  "<|13.70|>": 51049,
+  "<|13.72|>": 51050,
+  "<|13.74|>": 51051,
+  "<|13.76|>": 51052,
+  "<|13.78|>": 51053,
+  "<|13.80|>": 51054,
+  "<|13.82|>": 51055,
+  "<|13.84|>": 51056,
+  "<|13.86|>": 51057,
+  "<|13.88|>": 51058,
+  "<|13.90|>": 51059,
+  "<|13.92|>": 51060,
+  "<|13.94|>": 51061,
+  "<|13.96|>": 51062,
+  "<|13.98|>": 51063,
+  "<|14.00|>": 51064,
+  "<|14.02|>": 51065,
+  "<|14.04|>": 51066,
+  "<|14.06|>": 51067,
+  "<|14.08|>": 51068,
+  "<|14.10|>": 51069,
+  "<|14.12|>": 51070,
+  "<|14.14|>": 51071,
+  "<|14.16|>": 51072,
+  "<|14.18|>": 51073,
+  "<|14.20|>": 51074,
+  "<|14.22|>": 51075,
+  "<|14.24|>": 51076,
+  "<|14.26|>": 51077,
+  "<|14.28|>": 51078,
+  "<|14.30|>": 51079,
+  "<|14.32|>": 51080,
+  "<|14.34|>": 51081,
+  "<|14.36|>": 51082,
+  "<|14.38|>": 51083,
+  "<|14.40|>": 51084,
+  "<|14.42|>": 51085,
+  "<|14.44|>": 51086,
+  "<|14.46|>": 51087,
+  "<|14.48|>": 51088,
+  "<|14.50|>": 51089,
+  "<|14.52|>": 51090,
+  "<|14.54|>": 51091,
+  "<|14.56|>": 51092,
+  "<|14.58|>": 51093,
+  "<|14.60|>": 51094,
+  "<|14.62|>": 51095,
+  "<|14.64|>": 51096,
+  "<|14.66|>": 51097,
+  "<|14.68|>": 51098,
+  "<|14.70|>": 51099,
+  "<|14.72|>": 51100,
+  "<|14.74|>": 51101,
+  "<|14.76|>": 51102,
+  "<|14.78|>": 51103,
+  "<|14.80|>": 51104,
+  "<|14.82|>": 51105,
+  "<|14.84|>": 51106,
+  "<|14.86|>": 51107,
+  "<|14.88|>": 51108,
+  "<|14.90|>": 51109,
+  "<|14.92|>": 51110,
+  "<|14.94|>": 51111,
+  "<|14.96|>": 51112,
+  "<|14.98|>": 51113,
+  "<|15.00|>": 51114,
+  "<|15.02|>": 51115,
+  "<|15.04|>": 51116,
+  "<|15.06|>": 51117,
+  "<|15.08|>": 51118,
+  "<|15.10|>": 51119,
+  "<|15.12|>": 51120,
+  "<|15.14|>": 51121,
+  "<|15.16|>": 51122,
+  "<|15.18|>": 51123,
+  "<|15.20|>": 51124,
+  "<|15.22|>": 51125,
+  "<|15.24|>": 51126,
+  "<|15.26|>": 51127,
+  "<|15.28|>": 51128,
+  "<|15.30|>": 51129,
+  "<|15.32|>": 51130,
+  "<|15.34|>": 51131,
+  "<|15.36|>": 51132,
+  "<|15.38|>": 51133,
+  "<|15.40|>": 51134,
+  "<|15.42|>": 51135,
+  "<|15.44|>": 51136,
+  "<|15.46|>": 51137,
+  "<|15.48|>": 51138,
+  "<|15.50|>": 51139,
+  "<|15.52|>": 51140,
+  "<|15.54|>": 51141,
+  "<|15.56|>": 51142,
+  "<|15.58|>": 51143,
+  "<|15.60|>": 51144,
+  "<|15.62|>": 51145,
+  "<|15.64|>": 51146,
+  "<|15.66|>": 51147,
+  "<|15.68|>": 51148,
+  "<|15.70|>": 51149,
+  "<|15.72|>": 51150,
+  "<|15.74|>": 51151,
+  "<|15.76|>": 51152,
+  "<|15.78|>": 51153,
+  "<|15.80|>": 51154,
+  "<|15.82|>": 51155,
+  "<|15.84|>": 51156,
+  "<|15.86|>": 51157,
+  "<|15.88|>": 51158,
+  "<|15.90|>": 51159,
+  "<|15.92|>": 51160,
+  "<|15.94|>": 51161,
+  "<|15.96|>": 51162,
+  "<|15.98|>": 51163,
+  "<|16.00|>": 51164,
+  "<|16.02|>": 51165,
+  "<|16.04|>": 51166,
+  "<|16.06|>": 51167,
+  "<|16.08|>": 51168,
+  "<|16.10|>": 51169,
+  "<|16.12|>": 51170,
+  "<|16.14|>": 51171,
+  "<|16.16|>": 51172,
+  "<|16.18|>": 51173,
+  "<|16.20|>": 51174,
+  "<|16.22|>": 51175,
+  "<|16.24|>": 51176,
+  "<|16.26|>": 51177,
+  "<|16.28|>": 51178,
+  "<|16.30|>": 51179,
+  "<|16.32|>": 51180,
+  "<|16.34|>": 51181,
+  "<|16.36|>": 51182,
+  "<|16.38|>": 51183,
+  "<|16.40|>": 51184,
+  "<|16.42|>": 51185,
+  "<|16.44|>": 51186,
+  "<|16.46|>": 51187,
+  "<|16.48|>": 51188,
+  "<|16.50|>": 51189,
+  "<|16.52|>": 51190,
+  "<|16.54|>": 51191,
+  "<|16.56|>": 51192,
+  "<|16.58|>": 51193,
+  "<|16.60|>": 51194,
+  "<|16.62|>": 51195,
+  "<|16.64|>": 51196,
+  "<|16.66|>": 51197,
+  "<|16.68|>": 51198,
+  "<|16.70|>": 51199,
+  "<|16.72|>": 51200,
+  "<|16.74|>": 51201,
+  "<|16.76|>": 51202,
+  "<|16.78|>": 51203,
+  "<|16.80|>": 51204,
+  "<|16.82|>": 51205,
+  "<|16.84|>": 51206,
+  "<|16.86|>": 51207,
+  "<|16.88|>": 51208,
+  "<|16.90|>": 51209,
+  "<|16.92|>": 51210,
+  "<|16.94|>": 51211,
+  "<|16.96|>": 51212,
+  "<|16.98|>": 51213,
+  "<|17.00|>": 51214,
+  "<|17.02|>": 51215,
+  "<|17.04|>": 51216,
+  "<|17.06|>": 51217,
+  "<|17.08|>": 51218,
+  "<|17.10|>": 51219,
+  "<|17.12|>": 51220,
+  "<|17.14|>": 51221,
+  "<|17.16|>": 51222,
+  "<|17.18|>": 51223,
+  "<|17.20|>": 51224,
+  "<|17.22|>": 51225,
+  "<|17.24|>": 51226,
+  "<|17.26|>": 51227,
+  "<|17.28|>": 51228,
+  "<|17.30|>": 51229,
+  "<|17.32|>": 51230,
+  "<|17.34|>": 51231,
+  "<|17.36|>": 51232,
+  "<|17.38|>": 51233,
+  "<|17.40|>": 51234,
+  "<|17.42|>": 51235,
+  "<|17.44|>": 51236,
+  "<|17.46|>": 51237,
+  "<|17.48|>": 51238,
+  "<|17.50|>": 51239,
+  "<|17.52|>": 51240,
+  "<|17.54|>": 51241,
+  "<|17.56|>": 51242,
+  "<|17.58|>": 51243,
+  "<|17.60|>": 51244,
+  "<|17.62|>": 51245,
+  "<|17.64|>": 51246,
+  "<|17.66|>": 51247,
+  "<|17.68|>": 51248,
+  "<|17.70|>": 51249,
+  "<|17.72|>": 51250,
+  "<|17.74|>": 51251,
+  "<|17.76|>": 51252,
+  "<|17.78|>": 51253,
+  "<|17.80|>": 51254,
+  "<|17.82|>": 51255,
+  "<|17.84|>": 51256,
+  "<|17.86|>": 51257,
+  "<|17.88|>": 51258,
+  "<|17.90|>": 51259,
+  "<|17.92|>": 51260,
+  "<|17.94|>": 51261,
+  "<|17.96|>": 51262,
+  "<|17.98|>": 51263,
+  "<|18.00|>": 51264,
+  "<|18.02|>": 51265,
+  "<|18.04|>": 51266,
+  "<|18.06|>": 51267,
+  "<|18.08|>": 51268,
+  "<|18.10|>": 51269,
+  "<|18.12|>": 51270,
+  "<|18.14|>": 51271,
+  "<|18.16|>": 51272,
+  "<|18.18|>": 51273,
+  "<|18.20|>": 51274,
+  "<|18.22|>": 51275,
+  "<|18.24|>": 51276,
+  "<|18.26|>": 51277,
+  "<|18.28|>": 51278,
+  "<|18.30|>": 51279,
+  "<|18.32|>": 51280,
+  "<|18.34|>": 51281,
+  "<|18.36|>": 51282,
+  "<|18.38|>": 51283,
+  "<|18.40|>": 51284,
+  "<|18.42|>": 51285,
+  "<|18.44|>": 51286,
+  "<|18.46|>": 51287,
+  "<|18.48|>": 51288,
+  "<|18.50|>": 51289,
+  "<|18.52|>": 51290,
+  "<|18.54|>": 51291,
+  "<|18.56|>": 51292,
+  "<|18.58|>": 51293,
+  "<|18.60|>": 51294,
+  "<|18.62|>": 51295,
+  "<|18.64|>": 51296,
+  "<|18.66|>": 51297,
+  "<|18.68|>": 51298,
+  "<|18.70|>": 51299,
+  "<|18.72|>": 51300,
+  "<|18.74|>": 51301,
+  "<|18.76|>": 51302,
+  "<|18.78|>": 51303,
+  "<|18.80|>": 51304,
+  "<|18.82|>": 51305,
+  "<|18.84|>": 51306,
+  "<|18.86|>": 51307,
+  "<|18.88|>": 51308,
+  "<|18.90|>": 51309,
+  "<|18.92|>": 51310,
+  "<|18.94|>": 51311,
+  "<|18.96|>": 51312,
+  "<|18.98|>": 51313,
+  "<|19.00|>": 51314,
+  "<|19.02|>": 51315,
+  "<|19.04|>": 51316,
+  "<|19.06|>": 51317,
+  "<|19.08|>": 51318,
+  "<|19.10|>": 51319,
+  "<|19.12|>": 51320,
+  "<|19.14|>": 51321,
+  "<|19.16|>": 51322,
+  "<|19.18|>": 51323,
+  "<|19.20|>": 51324,
+  "<|19.22|>": 51325,
+  "<|19.24|>": 51326,
+  "<|19.26|>": 51327,
+  "<|19.28|>": 51328,
+  "<|19.30|>": 51329,
+  "<|19.32|>": 51330,
+  "<|19.34|>": 51331,
+  "<|19.36|>": 51332,
+  "<|19.38|>": 51333,
+  "<|19.40|>": 51334,
+  "<|19.42|>": 51335,
+  "<|19.44|>": 51336,
+  "<|19.46|>": 51337,
+  "<|19.48|>": 51338,
+  "<|19.50|>": 51339,
+  "<|19.52|>": 51340,
+  "<|19.54|>": 51341,
+  "<|19.56|>": 51342,
+  "<|19.58|>": 51343,
+  "<|19.60|>": 51344,
+  "<|19.62|>": 51345,
+  "<|19.64|>": 51346,
+  "<|19.66|>": 51347,
+  "<|19.68|>": 51348,
+  "<|19.70|>": 51349,
+  "<|19.72|>": 51350,
+  "<|19.74|>": 51351,
+  "<|19.76|>": 51352,
+  "<|19.78|>": 51353,
+  "<|19.80|>": 51354,
+  "<|19.82|>": 51355,
+  "<|19.84|>": 51356,
+  "<|19.86|>": 51357,
+  "<|19.88|>": 51358,
+  "<|19.90|>": 51359,
+  "<|19.92|>": 51360,
+  "<|19.94|>": 51361,
+  "<|19.96|>": 51362,
+  "<|19.98|>": 51363,
+  "<|2.00|>": 50464,
+  "<|2.02|>": 50465,
+  "<|2.04|>": 50466,
+  "<|2.06|>": 50467,
+  "<|2.08|>": 50468,
+  "<|2.10|>": 50469,
+  "<|2.12|>": 50470,
+  "<|2.14|>": 50471,
+  "<|2.16|>": 50472,
+  "<|2.18|>": 50473,
+  "<|2.20|>": 50474,
+  "<|2.22|>": 50475,
+  "<|2.24|>": 50476,
+  "<|2.26|>": 50477,
+  "<|2.28|>": 50478,
+  "<|2.30|>": 50479,
+  "<|2.32|>": 50480,
+  "<|2.34|>": 50481,
+  "<|2.36|>": 50482,
+  "<|2.38|>": 50483,
+  "<|2.40|>": 50484,
+  "<|2.42|>": 50485,
+  "<|2.44|>": 50486,
+  "<|2.46|>": 50487,
+  "<|2.48|>": 50488,
+  "<|2.50|>": 50489,
+  "<|2.52|>": 50490,
+  "<|2.54|>": 50491,
+  "<|2.56|>": 50492,
+  "<|2.58|>": 50493,
+  "<|2.60|>": 50494,
+  "<|2.62|>": 50495,
+  "<|2.64|>": 50496,
+  "<|2.66|>": 50497,
+  "<|2.68|>": 50498,
+  "<|2.70|>": 50499,
+  "<|2.72|>": 50500,
+  "<|2.74|>": 50501,
+  "<|2.76|>": 50502,
+  "<|2.78|>": 50503,
+  "<|2.80|>": 50504,
+  "<|2.82|>": 50505,
+  "<|2.84|>": 50506,
+  "<|2.86|>": 50507,
+  "<|2.88|>": 50508,
+  "<|2.90|>": 50509,
+  "<|2.92|>": 50510,
+  "<|2.94|>": 50511,
+  "<|2.96|>": 50512,
+  "<|2.98|>": 50513,
+  "<|20.00|>": 51364,
+  "<|20.02|>": 51365,
+  "<|20.04|>": 51366,
+  "<|20.06|>": 51367,
+  "<|20.08|>": 51368,
+  "<|20.10|>": 51369,
+  "<|20.12|>": 51370,
+  "<|20.14|>": 51371,
+  "<|20.16|>": 51372,
+  "<|20.18|>": 51373,
+  "<|20.20|>": 51374,
+  "<|20.22|>": 51375,
+  "<|20.24|>": 51376,
+  "<|20.26|>": 51377,
+  "<|20.28|>": 51378,
+  "<|20.30|>": 51379,
+  "<|20.32|>": 51380,
+  "<|20.34|>": 51381,
+  "<|20.36|>": 51382,
+  "<|20.38|>": 51383,
+  "<|20.40|>": 51384,
+  "<|20.42|>": 51385,
+  "<|20.44|>": 51386,
+  "<|20.46|>": 51387,
+  "<|20.48|>": 51388,
+  "<|20.50|>": 51389,
+  "<|20.52|>": 51390,
+  "<|20.54|>": 51391,
+  "<|20.56|>": 51392,
+  "<|20.58|>": 51393,
+  "<|20.60|>": 51394,
+  "<|20.62|>": 51395,
+  "<|20.64|>": 51396,
+  "<|20.66|>": 51397,
+  "<|20.68|>": 51398,
+  "<|20.70|>": 51399,
+  "<|20.72|>": 51400,
+  "<|20.74|>": 51401,
+  "<|20.76|>": 51402,
+  "<|20.78|>": 51403,
+  "<|20.80|>": 51404,
+  "<|20.82|>": 51405,
+  "<|20.84|>": 51406,
+  "<|20.86|>": 51407,
+  "<|20.88|>": 51408,
+  "<|20.90|>": 51409,
+  "<|20.92|>": 51410,
+  "<|20.94|>": 51411,
+  "<|20.96|>": 51412,
+  "<|20.98|>": 51413,
+  "<|21.00|>": 51414,
+  "<|21.02|>": 51415,
+  "<|21.04|>": 51416,
+  "<|21.06|>": 51417,
+  "<|21.08|>": 51418,
+  "<|21.10|>": 51419,
+  "<|21.12|>": 51420,
+  "<|21.14|>": 51421,
+  "<|21.16|>": 51422,
+  "<|21.18|>": 51423,
+  "<|21.20|>": 51424,
+  "<|21.22|>": 51425,
+  "<|21.24|>": 51426,
+  "<|21.26|>": 51427,
+  "<|21.28|>": 51428,
+  "<|21.30|>": 51429,
+  "<|21.32|>": 51430,
+  "<|21.34|>": 51431,
+  "<|21.36|>": 51432,
+  "<|21.38|>": 51433,
+  "<|21.40|>": 51434,
+  "<|21.42|>": 51435,
+  "<|21.44|>": 51436,
+  "<|21.46|>": 51437,
+  "<|21.48|>": 51438,
+  "<|21.50|>": 51439,
+  "<|21.52|>": 51440,
+  "<|21.54|>": 51441,
+  "<|21.56|>": 51442,
+  "<|21.58|>": 51443,
+  "<|21.60|>": 51444,
+  "<|21.62|>": 51445,
+  "<|21.64|>": 51446,
+  "<|21.66|>": 51447,
+  "<|21.68|>": 51448,
+  "<|21.70|>": 51449,
+  "<|21.72|>": 51450,
+  "<|21.74|>": 51451,
+  "<|21.76|>": 51452,
+  "<|21.78|>": 51453,
+  "<|21.80|>": 51454,
+  "<|21.82|>": 51455,
+  "<|21.84|>": 51456,
+  "<|21.86|>": 51457,
+  "<|21.88|>": 51458,
+  "<|21.90|>": 51459,
+  "<|21.92|>": 51460,
+  "<|21.94|>": 51461,
+  "<|21.96|>": 51462,
+  "<|21.98|>": 51463,
+  "<|22.00|>": 51464,
+  "<|22.02|>": 51465,
+  "<|22.04|>": 51466,
+  "<|22.06|>": 51467,
+  "<|22.08|>": 51468,
+  "<|22.10|>": 51469,
+  "<|22.12|>": 51470,
+  "<|22.14|>": 51471,
+  "<|22.16|>": 51472,
+  "<|22.18|>": 51473,
+  "<|22.20|>": 51474,
+  "<|22.22|>": 51475,
+  "<|22.24|>": 51476,
+  "<|22.26|>": 51477,
+  "<|22.28|>": 51478,
+  "<|22.30|>": 51479,
+  "<|22.32|>": 51480,
+  "<|22.34|>": 51481,
+  "<|22.36|>": 51482,
+  "<|22.38|>": 51483,
+  "<|22.40|>": 51484,
+  "<|22.42|>": 51485,
+  "<|22.44|>": 51486,
+  "<|22.46|>": 51487,
+  "<|22.48|>": 51488,
+  "<|22.50|>": 51489,
+  "<|22.52|>": 51490,
+  "<|22.54|>": 51491,
+  "<|22.56|>": 51492,
+  "<|22.58|>": 51493,
+  "<|22.60|>": 51494,
+  "<|22.62|>": 51495,
+  "<|22.64|>": 51496,
+  "<|22.66|>": 51497,
+  "<|22.68|>": 51498,
+  "<|22.70|>": 51499,
+  "<|22.72|>": 51500,
+  "<|22.74|>": 51501,
+  "<|22.76|>": 51502,
+  "<|22.78|>": 51503,
+  "<|22.80|>": 51504,
+  "<|22.82|>": 51505,
+  "<|22.84|>": 51506,
+  "<|22.86|>": 51507,
+  "<|22.88|>": 51508,
+  "<|22.90|>": 51509,
+  "<|22.92|>": 51510,
+  "<|22.94|>": 51511,
+  "<|22.96|>": 51512,
+  "<|22.98|>": 51513,
+  "<|23.00|>": 51514,
+  "<|23.02|>": 51515,
+  "<|23.04|>": 51516,
+  "<|23.06|>": 51517,
+  "<|23.08|>": 51518,
+  "<|23.10|>": 51519,
+  "<|23.12|>": 51520,
+  "<|23.14|>": 51521,
+  "<|23.16|>": 51522,
+  "<|23.18|>": 51523,
+  "<|23.20|>": 51524,
+  "<|23.22|>": 51525,
+  "<|23.24|>": 51526,
+  "<|23.26|>": 51527,
+  "<|23.28|>": 51528,
+  "<|23.30|>": 51529,
+  "<|23.32|>": 51530,
+  "<|23.34|>": 51531,
+  "<|23.36|>": 51532,
+  "<|23.38|>": 51533,
+  "<|23.40|>": 51534,
+  "<|23.42|>": 51535,
+  "<|23.44|>": 51536,
+  "<|23.46|>": 51537,
+  "<|23.48|>": 51538,
+  "<|23.50|>": 51539,
+  "<|23.52|>": 51540,
+  "<|23.54|>": 51541,
+  "<|23.56|>": 51542,
+  "<|23.58|>": 51543,
+  "<|23.60|>": 51544,
+  "<|23.62|>": 51545,
+  "<|23.64|>": 51546,
+  "<|23.66|>": 51547,
+  "<|23.68|>": 51548,
+  "<|23.70|>": 51549,
+  "<|23.72|>": 51550,
+  "<|23.74|>": 51551,
+  "<|23.76|>": 51552,
+  "<|23.78|>": 51553,
+  "<|23.80|>": 51554,
+  "<|23.82|>": 51555,
+  "<|23.84|>": 51556,
+  "<|23.86|>": 51557,
+  "<|23.88|>": 51558,
+  "<|23.90|>": 51559,
+  "<|23.92|>": 51560,
+  "<|23.94|>": 51561,
+  "<|23.96|>": 51562,
+  "<|23.98|>": 51563,
+  "<|24.00|>": 51564,
+  "<|24.02|>": 51565,
+  "<|24.04|>": 51566,
+  "<|24.06|>": 51567,
+  "<|24.08|>": 51568,
+  "<|24.10|>": 51569,
+  "<|24.12|>": 51570,
+  "<|24.14|>": 51571,
+  "<|24.16|>": 51572,
+  "<|24.18|>": 51573,
+  "<|24.20|>": 51574,
+  "<|24.22|>": 51575,
+  "<|24.24|>": 51576,
+  "<|24.26|>": 51577,
+  "<|24.28|>": 51578,
+  "<|24.30|>": 51579,
+  "<|24.32|>": 51580,
+  "<|24.34|>": 51581,
+  "<|24.36|>": 51582,
+  "<|24.38|>": 51583,
+  "<|24.40|>": 51584,
+  "<|24.42|>": 51585,
+  "<|24.44|>": 51586,
+  "<|24.46|>": 51587,
+  "<|24.48|>": 51588,
+  "<|24.50|>": 51589,
+  "<|24.52|>": 51590,
+  "<|24.54|>": 51591,
+  "<|24.56|>": 51592,
+  "<|24.58|>": 51593,
+  "<|24.60|>": 51594,
+  "<|24.62|>": 51595,
+  "<|24.64|>": 51596,
+  "<|24.66|>": 51597,
+  "<|24.68|>": 51598,
+  "<|24.70|>": 51599,
+  "<|24.72|>": 51600,
+  "<|24.74|>": 51601,
+  "<|24.76|>": 51602,
+  "<|24.78|>": 51603,
+  "<|24.80|>": 51604,
+  "<|24.82|>": 51605,
+  "<|24.84|>": 51606,
+  "<|24.86|>": 51607,
+  "<|24.88|>": 51608,
+  "<|24.90|>": 51609,
+  "<|24.92|>": 51610,
+  "<|24.94|>": 51611,
+  "<|24.96|>": 51612,
+  "<|24.98|>": 51613,
+  "<|25.00|>": 51614,
+  "<|25.02|>": 51615,
+  "<|25.04|>": 51616,
+  "<|25.06|>": 51617,
+  "<|25.08|>": 51618,
+  "<|25.10|>": 51619,
+  "<|25.12|>": 51620,
+  "<|25.14|>": 51621,
+  "<|25.16|>": 51622,
+  "<|25.18|>": 51623,
+  "<|25.20|>": 51624,
+  "<|25.22|>": 51625,
+  "<|25.24|>": 51626,
+  "<|25.26|>": 51627,
+  "<|25.28|>": 51628,
+  "<|25.30|>": 51629,
+  "<|25.32|>": 51630,
+  "<|25.34|>": 51631,
+  "<|25.36|>": 51632,
+  "<|25.38|>": 51633,
+  "<|25.40|>": 51634,
+  "<|25.42|>": 51635,
+  "<|25.44|>": 51636,
+  "<|25.46|>": 51637,
+  "<|25.48|>": 51638,
+  "<|25.50|>": 51639,
+  "<|25.52|>": 51640,
+  "<|25.54|>": 51641,
+  "<|25.56|>": 51642,
+  "<|25.58|>": 51643,
+  "<|25.60|>": 51644,
+  "<|25.62|>": 51645,
+  "<|25.64|>": 51646,
+  "<|25.66|>": 51647,
+  "<|25.68|>": 51648,
+  "<|25.70|>": 51649,
+  "<|25.72|>": 51650,
+  "<|25.74|>": 51651,
+  "<|25.76|>": 51652,
+  "<|25.78|>": 51653,
+  "<|25.80|>": 51654,
+  "<|25.82|>": 51655,
+  "<|25.84|>": 51656,
+  "<|25.86|>": 51657,
+  "<|25.88|>": 51658,
+  "<|25.90|>": 51659,
+  "<|25.92|>": 51660,
+  "<|25.94|>": 51661,
+  "<|25.96|>": 51662,
+  "<|25.98|>": 51663,
+  "<|26.00|>": 51664,
+  "<|26.02|>": 51665,
+  "<|26.04|>": 51666,
+  "<|26.06|>": 51667,
+  "<|26.08|>": 51668,
+  "<|26.10|>": 51669,
+  "<|26.12|>": 51670,
+  "<|26.14|>": 51671,
+  "<|26.16|>": 51672,
+  "<|26.18|>": 51673,
+  "<|26.20|>": 51674,
+  "<|26.22|>": 51675,
+  "<|26.24|>": 51676,
+  "<|26.26|>": 51677,
+  "<|26.28|>": 51678,
+  "<|26.30|>": 51679,
+  "<|26.32|>": 51680,
+  "<|26.34|>": 51681,
+  "<|26.36|>": 51682,
+  "<|26.38|>": 51683,
+  "<|26.40|>": 51684,
+  "<|26.42|>": 51685,
+  "<|26.44|>": 51686,
+  "<|26.46|>": 51687,
+  "<|26.48|>": 51688,
+  "<|26.50|>": 51689,
+  "<|26.52|>": 51690,
+  "<|26.54|>": 51691,
+  "<|26.56|>": 51692,
+  "<|26.58|>": 51693,
+  "<|26.60|>": 51694,
+  "<|26.62|>": 51695,
+  "<|26.64|>": 51696,
+  "<|26.66|>": 51697,
+  "<|26.68|>": 51698,
+  "<|26.70|>": 51699,
+  "<|26.72|>": 51700,
+  "<|26.74|>": 51701,
+  "<|26.76|>": 51702,
+  "<|26.78|>": 51703,
+  "<|26.80|>": 51704,
+  "<|26.82|>": 51705,
+  "<|26.84|>": 51706,
+  "<|26.86|>": 51707,
+  "<|26.88|>": 51708,
+  "<|26.90|>": 51709,
+  "<|26.92|>": 51710,
+  "<|26.94|>": 51711,
+  "<|26.96|>": 51712,
+  "<|26.98|>": 51713,
+  "<|27.00|>": 51714,
+  "<|27.02|>": 51715,
+  "<|27.04|>": 51716,
+  "<|27.06|>": 51717,
+  "<|27.08|>": 51718,
+  "<|27.10|>": 51719,
+  "<|27.12|>": 51720,
+  "<|27.14|>": 51721,
+  "<|27.16|>": 51722,
+  "<|27.18|>": 51723,
+  "<|27.20|>": 51724,
+  "<|27.22|>": 51725,
+  "<|27.24|>": 51726,
+  "<|27.26|>": 51727,
+  "<|27.28|>": 51728,
+  "<|27.30|>": 51729,
+  "<|27.32|>": 51730,
+  "<|27.34|>": 51731,
+  "<|27.36|>": 51732,
+  "<|27.38|>": 51733,
+  "<|27.40|>": 51734,
+  "<|27.42|>": 51735,
+  "<|27.44|>": 51736,
+  "<|27.46|>": 51737,
+  "<|27.48|>": 51738,
+  "<|27.50|>": 51739,
+  "<|27.52|>": 51740,
+  "<|27.54|>": 51741,
+  "<|27.56|>": 51742,
+  "<|27.58|>": 51743,
+  "<|27.60|>": 51744,
+  "<|27.62|>": 51745,
+  "<|27.64|>": 51746,
+  "<|27.66|>": 51747,
+  "<|27.68|>": 51748,
+  "<|27.70|>": 51749,
+  "<|27.72|>": 51750,
+  "<|27.74|>": 51751,
+  "<|27.76|>": 51752,
+  "<|27.78|>": 51753,
+  "<|27.80|>": 51754,
+  "<|27.82|>": 51755,
+  "<|27.84|>": 51756,
+  "<|27.86|>": 51757,
+  "<|27.88|>": 51758,
+  "<|27.90|>": 51759,
+  "<|27.92|>": 51760,
+  "<|27.94|>": 51761,
+  "<|27.96|>": 51762,
+  "<|27.98|>": 51763,
+  "<|28.00|>": 51764,
+  "<|28.02|>": 51765,
+  "<|28.04|>": 51766,
+  "<|28.06|>": 51767,
+  "<|28.08|>": 51768,
+  "<|28.10|>": 51769,
+  "<|28.12|>": 51770,
+  "<|28.14|>": 51771,
+  "<|28.16|>": 51772,
+  "<|28.18|>": 51773,
+  "<|28.20|>": 51774,
+  "<|28.22|>": 51775,
+  "<|28.24|>": 51776,
+  "<|28.26|>": 51777,
+  "<|28.28|>": 51778,
+  "<|28.30|>": 51779,
+  "<|28.32|>": 51780,
+  "<|28.34|>": 51781,
+  "<|28.36|>": 51782,
+  "<|28.38|>": 51783,
+  "<|28.40|>": 51784,
+  "<|28.42|>": 51785,
+  "<|28.44|>": 51786,
+  "<|28.46|>": 51787,
+  "<|28.48|>": 51788,
+  "<|28.50|>": 51789,
+  "<|28.52|>": 51790,
+  "<|28.54|>": 51791,
+  "<|28.56|>": 51792,
+  "<|28.58|>": 51793,
+  "<|28.60|>": 51794,
+  "<|28.62|>": 51795,
+  "<|28.64|>": 51796,
+  "<|28.66|>": 51797,
+  "<|28.68|>": 51798,
+  "<|28.70|>": 51799,
+  "<|28.72|>": 51800,
+  "<|28.74|>": 51801,
+  "<|28.76|>": 51802,
+  "<|28.78|>": 51803,
+  "<|28.80|>": 51804,
+  "<|28.82|>": 51805,
+  "<|28.84|>": 51806,
+  "<|28.86|>": 51807,
+  "<|28.88|>": 51808,
+  "<|28.90|>": 51809,
+  "<|28.92|>": 51810,
+  "<|28.94|>": 51811,
+  "<|28.96|>": 51812,
+  "<|28.98|>": 51813,
+  "<|29.00|>": 51814,
+  "<|29.02|>": 51815,
+  "<|29.04|>": 51816,
+  "<|29.06|>": 51817,
+  "<|29.08|>": 51818,
+  "<|29.10|>": 51819,
+  "<|29.12|>": 51820,
+  "<|29.14|>": 51821,
+  "<|29.16|>": 51822,
+  "<|29.18|>": 51823,
+  "<|29.20|>": 51824,
+  "<|29.22|>": 51825,
+  "<|29.24|>": 51826,
+  "<|29.26|>": 51827,
+  "<|29.28|>": 51828,
+  "<|29.30|>": 51829,
+  "<|29.32|>": 51830,
+  "<|29.34|>": 51831,
+  "<|29.36|>": 51832,
+  "<|29.38|>": 51833,
+  "<|29.40|>": 51834,
+  "<|29.42|>": 51835,
+  "<|29.44|>": 51836,
+  "<|29.46|>": 51837,
+  "<|29.48|>": 51838,
+  "<|29.50|>": 51839,
+  "<|29.52|>": 51840,
+  "<|29.54|>": 51841,
+  "<|29.56|>": 51842,
+  "<|29.58|>": 51843,
+  "<|29.60|>": 51844,
+  "<|29.62|>": 51845,
+  "<|29.64|>": 51846,
+  "<|29.66|>": 51847,
+  "<|29.68|>": 51848,
+  "<|29.70|>": 51849,
+  "<|29.72|>": 51850,
+  "<|29.74|>": 51851,
+  "<|29.76|>": 51852,
+  "<|29.78|>": 51853,
+  "<|29.80|>": 51854,
+  "<|29.82|>": 51855,
+  "<|29.84|>": 51856,
+  "<|29.86|>": 51857,
+  "<|29.88|>": 51858,
+  "<|29.90|>": 51859,
+  "<|29.92|>": 51860,
+  "<|29.94|>": 51861,
+  "<|29.96|>": 51862,
+  "<|29.98|>": 51863,
+  "<|3.00|>": 50514,
+  "<|3.02|>": 50515,
+  "<|3.04|>": 50516,
+  "<|3.06|>": 50517,
+  "<|3.08|>": 50518,
+  "<|3.10|>": 50519,
+  "<|3.12|>": 50520,
+  "<|3.14|>": 50521,
+  "<|3.16|>": 50522,
+  "<|3.18|>": 50523,
+  "<|3.20|>": 50524,
+  "<|3.22|>": 50525,
+  "<|3.24|>": 50526,
+  "<|3.26|>": 50527,
+  "<|3.28|>": 50528,
+  "<|3.30|>": 50529,
+  "<|3.32|>": 50530,
+  "<|3.34|>": 50531,
+  "<|3.36|>": 50532,
+  "<|3.38|>": 50533,
+  "<|3.40|>": 50534,
+  "<|3.42|>": 50535,
+  "<|3.44|>": 50536,
+  "<|3.46|>": 50537,
+  "<|3.48|>": 50538,
+  "<|3.50|>": 50539,
+  "<|3.52|>": 50540,
+  "<|3.54|>": 50541,
+  "<|3.56|>": 50542,
+  "<|3.58|>": 50543,
+  "<|3.60|>": 50544,
+  "<|3.62|>": 50545,
+  "<|3.64|>": 50546,
+  "<|3.66|>": 50547,
+  "<|3.68|>": 50548,
+  "<|3.70|>": 50549,
+  "<|3.72|>": 50550,
+  "<|3.74|>": 50551,
+  "<|3.76|>": 50552,
+  "<|3.78|>": 50553,
+  "<|3.80|>": 50554,
+  "<|3.82|>": 50555,
+  "<|3.84|>": 50556,
+  "<|3.86|>": 50557,
+  "<|3.88|>": 50558,
+  "<|3.90|>": 50559,
+  "<|3.92|>": 50560,
+  "<|3.94|>": 50561,
+  "<|3.96|>": 50562,
+  "<|3.98|>": 50563,
+  "<|30.00|>": 51864,
+  "<|4.00|>": 50564,
+  "<|4.02|>": 50565,
+  "<|4.04|>": 50566,
+  "<|4.06|>": 50567,
+  "<|4.08|>": 50568,
+  "<|4.10|>": 50569,
+  "<|4.12|>": 50570,
+  "<|4.14|>": 50571,
+  "<|4.16|>": 50572,
+  "<|4.18|>": 50573,
+  "<|4.20|>": 50574,
+  "<|4.22|>": 50575,
+  "<|4.24|>": 50576,
+  "<|4.26|>": 50577,
+  "<|4.28|>": 50578,
+  "<|4.30|>": 50579,
+  "<|4.32|>": 50580,
+  "<|4.34|>": 50581,
+  "<|4.36|>": 50582,
+  "<|4.38|>": 50583,
+  "<|4.40|>": 50584,
+  "<|4.42|>": 50585,
+  "<|4.44|>": 50586,
+  "<|4.46|>": 50587,
+  "<|4.48|>": 50588,
+  "<|4.50|>": 50589,
+  "<|4.52|>": 50590,
+  "<|4.54|>": 50591,
+  "<|4.56|>": 50592,
+  "<|4.58|>": 50593,
+  "<|4.60|>": 50594,
+  "<|4.62|>": 50595,
+  "<|4.64|>": 50596,
+  "<|4.66|>": 50597,
+  "<|4.68|>": 50598,
+  "<|4.70|>": 50599,
+  "<|4.72|>": 50600,
+  "<|4.74|>": 50601,
+  "<|4.76|>": 50602,
+  "<|4.78|>": 50603,
+  "<|4.80|>": 50604,
+  "<|4.82|>": 50605,
+  "<|4.84|>": 50606,
+  "<|4.86|>": 50607,
+  "<|4.88|>": 50608,
+  "<|4.90|>": 50609,
+  "<|4.92|>": 50610,
+  "<|4.94|>": 50611,
+  "<|4.96|>": 50612,
+  "<|4.98|>": 50613,
+  "<|5.00|>": 50614,
+  "<|5.02|>": 50615,
+  "<|5.04|>": 50616,
+  "<|5.06|>": 50617,
+  "<|5.08|>": 50618,
+  "<|5.10|>": 50619,
+  "<|5.12|>": 50620,
+  "<|5.14|>": 50621,
+  "<|5.16|>": 50622,
+  "<|5.18|>": 50623,
+  "<|5.20|>": 50624,
+  "<|5.22|>": 50625,
+  "<|5.24|>": 50626,
+  "<|5.26|>": 50627,
+  "<|5.28|>": 50628,
+  "<|5.30|>": 50629,
+  "<|5.32|>": 50630,
+  "<|5.34|>": 50631,
+  "<|5.36|>": 50632,
+  "<|5.38|>": 50633,
+  "<|5.40|>": 50634,
+  "<|5.42|>": 50635,
+  "<|5.44|>": 50636,
+  "<|5.46|>": 50637,
+  "<|5.48|>": 50638,
+  "<|5.50|>": 50639,
+  "<|5.52|>": 50640,
+  "<|5.54|>": 50641,
+  "<|5.56|>": 50642,
+  "<|5.58|>": 50643,
+  "<|5.60|>": 50644,
+  "<|5.62|>": 50645,
+  "<|5.64|>": 50646,
+  "<|5.66|>": 50647,
+  "<|5.68|>": 50648,
+  "<|5.70|>": 50649,
+  "<|5.72|>": 50650,
+  "<|5.74|>": 50651,
+  "<|5.76|>": 50652,
+  "<|5.78|>": 50653,
+  "<|5.80|>": 50654,
+  "<|5.82|>": 50655,
+  "<|5.84|>": 50656,
+  "<|5.86|>": 50657,
+  "<|5.88|>": 50658,
+  "<|5.90|>": 50659,
+  "<|5.92|>": 50660,
+  "<|5.94|>": 50661,
+  "<|5.96|>": 50662,
+  "<|5.98|>": 50663,
+  "<|6.00|>": 50664,
+  "<|6.02|>": 50665,
+  "<|6.04|>": 50666,
+  "<|6.06|>": 50667,
+  "<|6.08|>": 50668,
+  "<|6.10|>": 50669,
+  "<|6.12|>": 50670,
+  "<|6.14|>": 50671,
+  "<|6.16|>": 50672,
+  "<|6.18|>": 50673,
+  "<|6.20|>": 50674,
+  "<|6.22|>": 50675,
+  "<|6.24|>": 50676,
+  "<|6.26|>": 50677,
+  "<|6.28|>": 50678,
+  "<|6.30|>": 50679,
+  "<|6.32|>": 50680,
+  "<|6.34|>": 50681,
+  "<|6.36|>": 50682,
+  "<|6.38|>": 50683,
+  "<|6.40|>": 50684,
+  "<|6.42|>": 50685,
+  "<|6.44|>": 50686,
+  "<|6.46|>": 50687,
+  "<|6.48|>": 50688,
+  "<|6.50|>": 50689,
+  "<|6.52|>": 50690,
+  "<|6.54|>": 50691,
+  "<|6.56|>": 50692,
+  "<|6.58|>": 50693,
+  "<|6.60|>": 50694,
+  "<|6.62|>": 50695,
+  "<|6.64|>": 50696,
+  "<|6.66|>": 50697,
+  "<|6.68|>": 50698,
+  "<|6.70|>": 50699,
+  "<|6.72|>": 50700,
+  "<|6.74|>": 50701,
+  "<|6.76|>": 50702,
+  "<|6.78|>": 50703,
+  "<|6.80|>": 50704,
+  "<|6.82|>": 50705,
+  "<|6.84|>": 50706,
+  "<|6.86|>": 50707,
+  "<|6.88|>": 50708,
+  "<|6.90|>": 50709,
+  "<|6.92|>": 50710,
+  "<|6.94|>": 50711,
+  "<|6.96|>": 50712,
+  "<|6.98|>": 50713,
+  "<|7.00|>": 50714,
+  "<|7.02|>": 50715,
+  "<|7.04|>": 50716,
+  "<|7.06|>": 50717,
+  "<|7.08|>": 50718,
+  "<|7.10|>": 50719,
+  "<|7.12|>": 50720,
+  "<|7.14|>": 50721,
+  "<|7.16|>": 50722,
+  "<|7.18|>": 50723,
+  "<|7.20|>": 50724,
+  "<|7.22|>": 50725,
+  "<|7.24|>": 50726,
+  "<|7.26|>": 50727,
+  "<|7.28|>": 50728,
+  "<|7.30|>": 50729,
+  "<|7.32|>": 50730,
+  "<|7.34|>": 50731,
+  "<|7.36|>": 50732,
+  "<|7.38|>": 50733,
+  "<|7.40|>": 50734,
+  "<|7.42|>": 50735,
+  "<|7.44|>": 50736,
+  "<|7.46|>": 50737,
+  "<|7.48|>": 50738,
+  "<|7.50|>": 50739,
+  "<|7.52|>": 50740,
+  "<|7.54|>": 50741,
+  "<|7.56|>": 50742,
+  "<|7.58|>": 50743,
+  "<|7.60|>": 50744,
+  "<|7.62|>": 50745,
+  "<|7.64|>": 50746,
+  "<|7.66|>": 50747,
+  "<|7.68|>": 50748,
+  "<|7.70|>": 50749,
+  "<|7.72|>": 50750,
+  "<|7.74|>": 50751,
+  "<|7.76|>": 50752,
+  "<|7.78|>": 50753,
+  "<|7.80|>": 50754,
+  "<|7.82|>": 50755,
+  "<|7.84|>": 50756,
+  "<|7.86|>": 50757,
+  "<|7.88|>": 50758,
+  "<|7.90|>": 50759,
+  "<|7.92|>": 50760,
+  "<|7.94|>": 50761,
+  "<|7.96|>": 50762,
+  "<|7.98|>": 50763,
+  "<|8.00|>": 50764,
+  "<|8.02|>": 50765,
+  "<|8.04|>": 50766,
+  "<|8.06|>": 50767,
+  "<|8.08|>": 50768,
+  "<|8.10|>": 50769,
+  "<|8.12|>": 50770,
+  "<|8.14|>": 50771,
+  "<|8.16|>": 50772,
+  "<|8.18|>": 50773,
+  "<|8.20|>": 50774,
+  "<|8.22|>": 50775,
+  "<|8.24|>": 50776,
+  "<|8.26|>": 50777,
+  "<|8.28|>": 50778,
+  "<|8.30|>": 50779,
+  "<|8.32|>": 50780,
+  "<|8.34|>": 50781,
+  "<|8.36|>": 50782,
+  "<|8.38|>": 50783,
+  "<|8.40|>": 50784,
+  "<|8.42|>": 50785,
+  "<|8.44|>": 50786,
+  "<|8.46|>": 50787,
+  "<|8.48|>": 50788,
+  "<|8.50|>": 50789,
+  "<|8.52|>": 50790,
+  "<|8.54|>": 50791,
+  "<|8.56|>": 50792,
+  "<|8.58|>": 50793,
+  "<|8.60|>": 50794,
+  "<|8.62|>": 50795,
+  "<|8.64|>": 50796,
+  "<|8.66|>": 50797,
+  "<|8.68|>": 50798,
+  "<|8.70|>": 50799,
+  "<|8.72|>": 50800,
+  "<|8.74|>": 50801,
+  "<|8.76|>": 50802,
+  "<|8.78|>": 50803,
+  "<|8.80|>": 50804,
+  "<|8.82|>": 50805,
+  "<|8.84|>": 50806,
+  "<|8.86|>": 50807,
+  "<|8.88|>": 50808,
+  "<|8.90|>": 50809,
+  "<|8.92|>": 50810,
+  "<|8.94|>": 50811,
+  "<|8.96|>": 50812,
+  "<|8.98|>": 50813,
+  "<|9.00|>": 50814,
+  "<|9.02|>": 50815,
+  "<|9.04|>": 50816,
+  "<|9.06|>": 50817,
+  "<|9.08|>": 50818,
+  "<|9.10|>": 50819,
+  "<|9.12|>": 50820,
+  "<|9.14|>": 50821,
+  "<|9.16|>": 50822,
+  "<|9.18|>": 50823,
+  "<|9.20|>": 50824,
+  "<|9.22|>": 50825,
+  "<|9.24|>": 50826,
+  "<|9.26|>": 50827,
+  "<|9.28|>": 50828,
+  "<|9.30|>": 50829,
+  "<|9.32|>": 50830,
+  "<|9.34|>": 50831,
+  "<|9.36|>": 50832,
+  "<|9.38|>": 50833,
+  "<|9.40|>": 50834,
+  "<|9.42|>": 50835,
+  "<|9.44|>": 50836,
+  "<|9.46|>": 50837,
+  "<|9.48|>": 50838,
+  "<|9.50|>": 50839,
+  "<|9.52|>": 50840,
+  "<|9.54|>": 50841,
+  "<|9.56|>": 50842,
+  "<|9.58|>": 50843,
+  "<|9.60|>": 50844,
+  "<|9.62|>": 50845,
+  "<|9.64|>": 50846,
+  "<|9.66|>": 50847,
+  "<|9.68|>": 50848,
+  "<|9.70|>": 50849,
+  "<|9.72|>": 50850,
+  "<|9.74|>": 50851,
+  "<|9.76|>": 50852,
+  "<|9.78|>": 50853,
+  "<|9.80|>": 50854,
+  "<|9.82|>": 50855,
+  "<|9.84|>": 50856,
+  "<|9.86|>": 50857,
+  "<|9.88|>": 50858,
+  "<|9.90|>": 50859,
+  "<|9.92|>": 50860,
+  "<|9.94|>": 50861,
+  "<|9.96|>": 50862,
+  "<|9.98|>": 50863,
+  "<|af|>": 50327,
+  "<|am|>": 50334,
+  "<|ar|>": 50272,
+  "<|as|>": 50350,
+  "<|az|>": 50304,
+  "<|ba|>": 50355,
+  "<|be|>": 50330,
+  "<|bg|>": 50292,
+  "<|bn|>": 50302,
+  "<|bo|>": 50347,
+  "<|br|>": 50309,
+  "<|bs|>": 50315,
+  "<|ca|>": 50270,
+  "<|cs|>": 50283,
+  "<|cy|>": 50297,
+  "<|da|>": 50285,
+  "<|de|>": 50261,
+  "<|el|>": 50281,
+  "<|en|>": 50259,
+  "<|es|>": 50262,
+  "<|et|>": 50307,
+  "<|eu|>": 50310,
+  "<|fa|>": 50300,
+  "<|fi|>": 50277,
+  "<|fo|>": 50338,
+  "<|fr|>": 50265,
+  "<|gl|>": 50319,
+  "<|gu|>": 50333,
+  "<|haw|>": 50352,
+  "<|ha|>": 50354,
+  "<|he|>": 50279,
+  "<|hi|>": 50276,
+  "<|hr|>": 50291,
+  "<|ht|>": 50339,
+  "<|hu|>": 50286,
+  "<|hy|>": 50312,
+  "<|id|>": 50275,
+  "<|is|>": 50311,
+  "<|it|>": 50274,
+  "<|ja|>": 50266,
+  "<|jw|>": 50356,
+  "<|ka|>": 50329,
+  "<|kk|>": 50316,
+  "<|km|>": 50323,
+  "<|kn|>": 50306,
+  "<|ko|>": 50264,
+  "<|la|>": 50294,
+  "<|lb|>": 50345,
+  "<|ln|>": 50353,
+  "<|lo|>": 50336,
+  "<|lt|>": 50293,
+  "<|lv|>": 50301,
+  "<|mg|>": 50349,
+  "<|mi|>": 50295,
+  "<|mk|>": 50308,
+  "<|ml|>": 50296,
+  "<|mn|>": 50314,
+  "<|mr|>": 50320,
+  "<|ms|>": 50282,
+  "<|mt|>": 50343,
+  "<|my|>": 50346,
+  "<|ne|>": 50313,
+  "<|nl|>": 50271,
+  "<|nn|>": 50342,
+  "<|nocaptions|>": 50362,
+  "<|notimestamps|>": 50363,
+  "<|no|>": 50288,
+  "<|oc|>": 50328,
+  "<|pa|>": 50321,
+  "<|pl|>": 50269,
+  "<|ps|>": 50340,
+  "<|pt|>": 50267,
+  "<|ro|>": 50284,
+  "<|ru|>": 50263,
+  "<|sa|>": 50344,
+  "<|sd|>": 50332,
+  "<|si|>": 50322,
+  "<|sk|>": 50298,
+  "<|sl|>": 50305,
+  "<|sn|>": 50324,
+  "<|so|>": 50326,
+  "<|sq|>": 50317,
+  "<|sr|>": 50303,
+  "<|startoflm|>": 50360,
+  "<|startofprev|>": 50361,
+  "<|startoftranscript|>": 50258,
+  "<|su|>": 50357,
+  "<|sv|>": 50273,
+  "<|sw|>": 50318,
+  "<|ta|>": 50287,
+  "<|te|>": 50299,
+  "<|tg|>": 50331,
+  "<|th|>": 50289,
+  "<|tk|>": 50341,
+  "<|tl|>": 50348,
+  "<|transcribe|>": 50359,
+  "<|translate|>": 50358,
+  "<|tr|>": 50268,
+  "<|tt|>": 50351,
+  "<|uk|>": 50280,
+  "<|ur|>": 50290,
+  "<|uz|>": 50337,
+  "<|vi|>": 50278,
+  "<|yi|>": 50335,
+  "<|yo|>": 50325,
+  "<|zh|>": 50260
+}

config.json

@@ -0,0 +1,57 @@
+{
+  "_name_or_path": "openai/whisper-base",
+  "activation_dropout": 0.0,
+  "activation_function": "gelu",
+  "apply_spec_augment": false,
+  "architectures": [
+    "WhisperMLAForConditionalGeneration"
+  ],
+  "attention_dropout": 0.0,
+  "begin_suppress_tokens": null,
+  "bos_token_id": 50257,
+  "classifier_proj_size": 256,
+  "d_model": 512,
+  "decoder_attention_heads": 8,
+  "decoder_ffn_dim": 2048,
+  "decoder_layerdrop": 0.0,
+  "decoder_layers": 6,
+  "decoder_start_token_id": 50258,
+  "dropout": 0.0,
+  "encoder_attention_heads": 8,
+  "encoder_ffn_dim": 2048,
+  "encoder_layerdrop": 0.0,
+  "encoder_layers": 6,
+  "eos_token_id": 50257,
+  "init_std": 0.02,
+  "is_encoder_decoder": true,
+  "mask_feature_length": 10,
+  "mask_feature_min_masks": 0,
+  "mask_feature_prob": 0.0,
+  "mask_time_length": 10,
+  "mask_time_min_masks": 2,
+  "mask_time_prob": 0.05,
+  "max_length": null,
+  "max_source_positions": 1500,
+  "max_target_positions": 448,
+  "median_filter_width": 7,
+  "model_type": "whisper_mla",
+  "num_hidden_layers": 6,
+  "num_mel_bins": 80,
+  "pad_token_id": 50257,
+  "scale_embedding": false,
+  "torch_dtype": "float32",
+  "transformers_version": "4.46.3",
+  "use_cache": true,
+  "use_weighted_layer_sum": false,
+  "vocab_size": 51865,
+  "auto_map": {
+    "AutoConfig": "configuration_whisper_mla.WhisperMLAConfig",
+    "AutoModel": "modeling_whisper_mla.WhisperMLAForConditionalGeneration",
+    "AutoModelForSpeechSeq2Seq": "modeling_whisper_mla.WhisperMLAForConditionalGeneration"
+  },
+  "mla_scheme": "partial_preservation",
+  "mla_scope": "decoder_self",
+  "mla_selection": "uniform",
+  "mla_low_rank": 16,
+  "mla_rope_dim": 32
+}

configuration_whisper_mla.py

@@ -0,0 +1,19 @@
+"""WhisperMLA configuration — WhisperConfig + the MLA conversion axes.
+
+Shipped inside released model repos (trust_remote_code); loaded by AutoConfig via auto_map.
+"""
+from transformers import WhisperConfig
+
+
+class WhisperMLAConfig(WhisperConfig):
+    model_type = "whisper_mla"
+
+    def __init__(self, mla_scheme: str = "partial_preservation", mla_scope: str = "decoder_self",
+                 mla_selection: str = "uniform", mla_low_rank: int = 16, mla_rope_dim: int = 32,
+                 **kwargs):
+        self.mla_scheme = mla_scheme
+        self.mla_scope = mla_scope
+        self.mla_selection = mla_selection
+        self.mla_low_rank = mla_low_rank
+        self.mla_rope_dim = mla_rope_dim
+        super().__init__(**kwargs)

generation_config.json

@@ -0,0 +1,246 @@
+{
+  "alignment_heads": [
+    [
+      3,
+      1
+    ],
+    [
+      4,
+      2
+    ],
+    [
+      4,
+      3
+    ],
+    [
+      4,
+      7
+    ],
+    [
+      5,
+      1
+    ],
+    [
+      5,
+      2
+    ],
+    [
+      5,
+      4
+    ],
+    [
+      5,
+      6
+    ]
+  ],
+  "begin_suppress_tokens": [
+    220,
+    50257
+  ],
+  "bos_token_id": 50257,
+  "decoder_start_token_id": 50258,
+  "eos_token_id": 50257,
+  "is_multilingual": true,
+  "lang_to_id": {
+    "<|af|>": 50327,
+    "<|am|>": 50334,
+    "<|ar|>": 50272,
+    "<|as|>": 50350,
+    "<|az|>": 50304,
+    "<|ba|>": 50355,
+    "<|be|>": 50330,
+    "<|bg|>": 50292,
+    "<|bn|>": 50302,
+    "<|bo|>": 50347,
+    "<|br|>": 50309,
+    "<|bs|>": 50315,
+    "<|ca|>": 50270,
+    "<|cs|>": 50283,
+    "<|cy|>": 50297,
+    "<|da|>": 50285,
+    "<|de|>": 50261,
+    "<|el|>": 50281,
+    "<|en|>": 50259,
+    "<|es|>": 50262,
+    "<|et|>": 50307,
+    "<|eu|>": 50310,
+    "<|fa|>": 50300,
+    "<|fi|>": 50277,
+    "<|fo|>": 50338,
+    "<|fr|>": 50265,
+    "<|gl|>": 50319,
+    "<|gu|>": 50333,
+    "<|haw|>": 50352,
+    "<|ha|>": 50354,
+    "<|he|>": 50279,
+    "<|hi|>": 50276,
+    "<|hr|>": 50291,
+    "<|ht|>": 50339,
+    "<|hu|>": 50286,
+    "<|hy|>": 50312,
+    "<|id|>": 50275,
+    "<|is|>": 50311,
+    "<|it|>": 50274,
+    "<|ja|>": 50266,
+    "<|jw|>": 50356,
+    "<|ka|>": 50329,
+    "<|kk|>": 50316,
+    "<|km|>": 50323,
+    "<|kn|>": 50306,
+    "<|ko|>": 50264,
+    "<|la|>": 50294,
+    "<|lb|>": 50345,
+    "<|ln|>": 50353,
+    "<|lo|>": 50336,
+    "<|lt|>": 50293,
+    "<|lv|>": 50301,
+    "<|mg|>": 50349,
+    "<|mi|>": 50295,
+    "<|mk|>": 50308,
+    "<|ml|>": 50296,
+    "<|mn|>": 50314,
+    "<|mr|>": 50320,
+    "<|ms|>": 50282,
+    "<|mt|>": 50343,
+    "<|my|>": 50346,
+    "<|ne|>": 50313,
+    "<|nl|>": 50271,
+    "<|nn|>": 50342,
+    "<|no|>": 50288,
+    "<|oc|>": 50328,
+    "<|pa|>": 50321,
+    "<|pl|>": 50269,
+    "<|ps|>": 50340,
+    "<|pt|>": 50267,
+    "<|ro|>": 50284,
+    "<|ru|>": 50263,
+    "<|sa|>": 50344,
+    "<|sd|>": 50332,
+    "<|si|>": 50322,
+    "<|sk|>": 50298,
+    "<|sl|>": 50305,
+    "<|sn|>": 50324,
+    "<|so|>": 50326,
+    "<|sq|>": 50317,
+    "<|sr|>": 50303,
+    "<|su|>": 50357,
+    "<|sv|>": 50273,
+    "<|sw|>": 50318,
+    "<|ta|>": 50287,
+    "<|te|>": 50299,
+    "<|tg|>": 50331,
+    "<|th|>": 50289,
+    "<|tk|>": 50341,
+    "<|tl|>": 50348,
+    "<|tr|>": 50268,
+    "<|tt|>": 50351,
+    "<|uk|>": 50280,
+    "<|ur|>": 50290,
+    "<|uz|>": 50337,
+    "<|vi|>": 50278,
+    "<|yi|>": 50335,
+    "<|yo|>": 50325,
+    "<|zh|>": 50260
+  },
+  "max_initial_timestamp_index": 50,
+  "max_length": 448,
+  "no_timestamps_token_id": 50363,
+  "pad_token_id": 50257,
+  "prev_sot_token_id": 50361,
+  "return_timestamps": false,
+  "suppress_tokens": [
+    1,
+    2,
+    7,
+    8,
+    9,
+    10,
+    14,
+    25,
+    26,
+    27,
+    28,
+    29,
+    31,
+    58,
+    59,
+    60,
+    61,
+    62,
+    63,
+    90,
+    91,
+    92,
+    93,
+    359,
+    503,
+    522,
+    542,
+    873,
+    893,
+    902,
+    918,
+    922,
+    931,
+    1350,
+    1853,
+    1982,
+    2460,
+    2627,
+    3246,
+    3253,
+    3268,
+    3536,
+    3846,
+    3961,
+    4183,
+    4667,
+    6585,
+    6647,
+    7273,
+    9061,
+    9383,
+    10428,
+    10929,
+    11938,
+    12033,
+    12331,
+    12562,
+    13793,
+    14157,
+    14635,
+    15265,
+    15618,
+    16553,
+    16604,
+    18362,
+    18956,
+    20075,
+    21675,
+    22520,
+    26130,
+    26161,
+    26435,
+    28279,
+    29464,
+    31650,
+    32302,
+    32470,
+    36865,
+    42863,
+    47425,
+    49870,
+    50254,
+    50258,
+    50358,
+    50359,
+    50360,
+    50361,
+    50362
+  ],
+  "task_to_id": {
+    "transcribe": 50359,
+    "translate": 50358
+  },
+  "transformers_version": "4.46.3"
+}

model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:be0ef903f6dda7ea68173e3d374fc1611f15abaee46af105f7663a3b025dbf21
+size 142468680

modeling_whisper.py

@@ -0,0 +1,1777 @@
+# This script is based on https://github.com/huggingface/transformers/blob/v4.29.1/src/transformers/models/whisper/modeling_whisper.py
+
+""" PyTorch Whisper model."""
+
+import math
+import random
+from typing import Optional, Tuple, Union
+
+import numpy as np
+import torch
+import torch.utils.checkpoint
+from torch import nn
+from torch.nn import CrossEntropyLoss
+
+from transformers.activations import ACT2FN
+from transformers.generation.logits_process import WhisperTimeStampLogitsProcessor
+from transformers.modeling_outputs import (
+    BaseModelOutput,
+    BaseModelOutputWithPastAndCrossAttentions,
+    Seq2SeqLMOutput,
+    Seq2SeqModelOutput,
+    SequenceClassifierOutput,
+)
+from transformers.modeling_utils import PreTrainedModel
+from transformers.utils import add_start_docstrings, add_start_docstrings_to_model_forward, logging, replace_return_docstrings
+from transformers.models.whisper.configuration_whisper import WhisperConfig
+from transformers.models.whisper.tokenization_whisper import TASK_IDS, TO_LANGUAGE_CODE
+
+
+logger = logging.get_logger(__name__)
+
+_CONFIG_FOR_DOC = "WhisperConfig"
+_CHECKPOINT_FOR_DOC = "openai/whisper-tiny"
+
+
+WHISPER_PRETRAINED_MODEL_ARCHIVE_LIST = [
+    "openai/whisper-base",
+    # See all Whisper models at https://huggingface.co/models?filter=whisper
+]
+
+
+# Copied from transformers.models.bart.modeling_bart.shift_tokens_right
+def shift_tokens_right(input_ids: torch.Tensor, pad_token_id: int, decoder_start_token_id: int):
+    """
+    Shift input ids one token to the right.
+    """
+    shifted_input_ids = input_ids.new_zeros(input_ids.shape)
+    shifted_input_ids[:, 1:] = input_ids[:, :-1].clone()
+    shifted_input_ids[:, 0] = decoder_start_token_id
+
+    if pad_token_id is None:
+        raise ValueError("self.model.config.pad_token_id has to be defined.")
+    # replace possible -100 values in labels by `pad_token_id`
+    shifted_input_ids.masked_fill_(shifted_input_ids == -100, pad_token_id)
+
+    return shifted_input_ids
+
+
+# Copied from transformers.models.bart.modeling_bart._make_causal_mask
+def _make_causal_mask(
+    input_ids_shape: torch.Size, dtype: torch.dtype, device: torch.device, past_key_values_length: int = 0
+):
+    """
+    Make causal mask used for bi-directional self-attention.
+    """
+    bsz, tgt_len = input_ids_shape
+    mask = torch.full((tgt_len, tgt_len), torch.tensor(torch.finfo(dtype).min, device=device), device=device)
+    mask_cond = torch.arange(mask.size(-1), device=device)
+    mask.masked_fill_(mask_cond < (mask_cond + 1).view(mask.size(-1), 1), 0)
+    mask = mask.to(dtype)
+
+    if past_key_values_length > 0:
+        mask = torch.cat([torch.zeros(tgt_len, past_key_values_length, dtype=dtype, device=device), mask], dim=-1)
+    return mask[None, None, :, :].expand(bsz, 1, tgt_len, tgt_len + past_key_values_length)
+
+
+# Copied from transformers.models.bart.modeling_bart._expand_mask
+def _expand_mask(mask: torch.Tensor, dtype: torch.dtype, tgt_len: Optional[int] = None):
+    """
+    Expands attention_mask from `[bsz, seq_len]` to `[bsz, 1, tgt_seq_len, src_seq_len]`.
+    """
+    bsz, src_len = mask.size()
+    tgt_len = tgt_len if tgt_len is not None else src_len
+
+    expanded_mask = mask[:, None, None, :].expand(bsz, 1, tgt_len, src_len).to(dtype)
+
+    inverted_mask = 1.0 - expanded_mask
+
+    return inverted_mask.masked_fill(inverted_mask.to(torch.bool), torch.finfo(dtype).min)
+
+
+# Copied from transformers.models.wav2vec2.modeling_wav2vec2._compute_mask_indices
+def _compute_mask_indices(
+    shape: Tuple[int, int],
+    mask_prob: float,
+    mask_length: int,
+    attention_mask: Optional[torch.LongTensor] = None,
+    min_masks: int = 0,
+) -> np.ndarray:
+    """
+    Computes random mask spans for a given shape. Used to implement [SpecAugment: A Simple Data Augmentation Method for
+    ASR](https://arxiv.org/abs/1904.08779). Note that this method is not optimized to run on TPU and should be run on
+    CPU as part of the preprocessing during training.
+
+    Args:
+        shape: The shape for which to compute masks. This should be of a tuple of size 2 where
+               the first element is the batch size and the second element is the length of the axis to span.
+        mask_prob:  The percentage of the whole axis (between 0 and 1) which will be masked. The number of
+                    independently generated mask spans of length `mask_length` is computed by
+                    `mask_prob*shape[1]/mask_length`. Note that due to overlaps, `mask_prob` is an upper bound and the
+                    actual percentage will be smaller.
+        mask_length: size of the mask
+        min_masks: minimum number of masked spans
+        attention_mask: A (right-padded) attention mask which independently shortens the feature axis of
+                        each batch dimension.
+    """
+    batch_size, sequence_length = shape
+
+    if mask_length < 1:
+        raise ValueError("`mask_length` has to be bigger than 0.")
+
+    if mask_length > sequence_length:
+        raise ValueError(
+            f"`mask_length` has to be smaller than `sequence_length`, but got `mask_length`: {mask_length}"
+            f" and `sequence_length`: {sequence_length}`"
+        )
+
+    # epsilon is used for probabilistic rounding
+    epsilon = np.random.rand(1).item()
+
+    def compute_num_masked_span(input_length):
+        """Given input length, compute how many spans should be masked"""
+        num_masked_span = int(mask_prob * input_length / mask_length + epsilon)
+        num_masked_span = max(num_masked_span, min_masks)
+
+        # make sure num masked span <= sequence_length
+        if num_masked_span * mask_length > sequence_length:
+            num_masked_span = sequence_length // mask_length
+
+        # make sure num_masked span is also <= input_length - (mask_length - 1)
+        if input_length - (mask_length - 1) < num_masked_span:
+            num_masked_span = max(input_length - (mask_length - 1), 0)
+
+        return num_masked_span
+
+    # compute number of masked spans in batch
+    input_lengths = (
+        attention_mask.sum(-1).detach().tolist()
+        if attention_mask is not None
+        else [sequence_length for _ in range(batch_size)]
+    )
+
+    # SpecAugment mask to fill
+    spec_aug_mask = np.zeros((batch_size, sequence_length), dtype=bool)
+    spec_aug_mask_idxs = []
+
+    max_num_masked_span = compute_num_masked_span(sequence_length)
+
+    if max_num_masked_span == 0:
+        return spec_aug_mask
+
+    for input_length in input_lengths:
+        # compute num of masked spans for this input
+        num_masked_span = compute_num_masked_span(input_length)
+
+        # get random indices to mask
+        spec_aug_mask_idx = np.random.choice(
+            np.arange(input_length - (mask_length - 1)), num_masked_span, replace=False
+        )
+
+        # pick first sampled index that will serve as a dummy index to pad vector
+        # to ensure same dimension for all batches due to probabilistic rounding
+        # Picking first sample just pads those vectors twice.
+        if len(spec_aug_mask_idx) == 0:
+            # this case can only happen if `input_length` is strictly smaller then
+            # `sequence_length` in which case the last token has to be a padding
+            # token which we can use as a dummy mask id
+            dummy_mask_idx = sequence_length - 1
+        else:
+            dummy_mask_idx = spec_aug_mask_idx[0]
+
+        spec_aug_mask_idx = np.concatenate(
+            [spec_aug_mask_idx, np.ones(max_num_masked_span - num_masked_span, dtype=np.int32) * dummy_mask_idx]
+        )
+        spec_aug_mask_idxs.append(spec_aug_mask_idx)
+
+    spec_aug_mask_idxs = np.array(spec_aug_mask_idxs)
+
+    # expand masked indices to masked spans
+    spec_aug_mask_idxs = np.broadcast_to(
+        spec_aug_mask_idxs[:, :, None], (batch_size, max_num_masked_span, mask_length)
+    )
+    spec_aug_mask_idxs = spec_aug_mask_idxs.reshape(batch_size, max_num_masked_span * mask_length)
+
+    # add offset to the starting indexes so that indexes now create a span
+    offsets = np.arange(mask_length)[None, None, :]
+    offsets = np.broadcast_to(offsets, (batch_size, max_num_masked_span, mask_length)).reshape(
+        batch_size, max_num_masked_span * mask_length
+    )
+    spec_aug_mask_idxs = spec_aug_mask_idxs + offsets
+
+    # ensure that we cannot have indices larger than sequence_length
+    if spec_aug_mask_idxs.max() > sequence_length - 1:
+        spec_aug_mask_idxs[spec_aug_mask_idxs > sequence_length - 1] = sequence_length - 1
+
+    # scatter indices to mask
+    np.put_along_axis(spec_aug_mask, spec_aug_mask_idxs, 1, -1)
+
+    return spec_aug_mask
+
+
+class WhisperPositionalEmbedding(nn.Embedding):
+    def __init__(self, num_positions: int, embedding_dim: int, padding_idx: Optional[int] = None):
+        super().__init__(num_positions, embedding_dim)
+
+    def forward(self, input_ids, past_key_values_length=0):
+        return self.weight[past_key_values_length : past_key_values_length + input_ids.shape[1]]
+
+
+class WhisperAttention(nn.Module):
+    """Multi-headed attention from 'Attention Is All You Need' paper"""
+
+    def __init__(
+        self,
+        embed_dim: int,
+        num_heads: int,
+        dropout: float = 0.0,
+        is_decoder: bool = False,
+        bias: bool = True,
+    ):
+        super().__init__()
+        self.embed_dim = embed_dim
+        self.num_heads = num_heads
+        self.dropout = dropout
+        self.head_dim = embed_dim // num_heads
+
+        if (self.head_dim * num_heads) != self.embed_dim:
+            raise ValueError(
+                f"embed_dim must be divisible by num_heads (got `embed_dim`: {self.embed_dim}"
+                f" and `num_heads`: {num_heads})."
+            )
+        self.scaling = self.head_dim**-0.5
+        self.is_decoder = is_decoder
+
+        self.k_proj = nn.Linear(embed_dim, embed_dim, bias=False)
+        self.v_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
+        self.q_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
+        self.out_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
+
+    # Copied from transformers.models.bart.modeling_bart.BartAttention._shape with BART->whisper
+    def _shape(self, tensor: torch.Tensor, seq_len: int, bsz: int):
+        return tensor.view(bsz, seq_len, self.num_heads, self.head_dim).transpose(1, 2).contiguous()
+
+    # Copied from transformers.models.bart.modeling_bart.BartAttention.forward with BART->whisper
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        key_value_states: Optional[torch.Tensor] = None,
+        past_key_value: Optional[Tuple[torch.Tensor]] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        layer_head_mask: Optional[torch.Tensor] = None,
+        output_attentions: bool = False,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+        """Input shape: Batch x Time x Channel"""
+
+        # if key_value_states are provided this layer is used as a cross-attention layer
+        # for the decoder
+        is_cross_attention = key_value_states is not None
+
+        bsz, tgt_len, _ = hidden_states.size()
+
+        # get query proj
+        query_states = self.q_proj(hidden_states) * self.scaling
+        # get key, value proj
+        # `past_key_value[0].shape[2] == key_value_states.shape[1]`
+        # is checking that the `sequence_length` of the `past_key_value` is the same as
+        # the provided `key_value_states` to support prefix tuning
+        if (
+            is_cross_attention
+            and past_key_value is not None
+            and past_key_value[0].shape[2] == key_value_states.shape[1]
+        ):
+            # reuse k,v, cross_attentions
+            key_states = past_key_value[0]
+            value_states = past_key_value[1]
+        elif is_cross_attention:
+            # cross_attentions
+            key_states = self._shape(self.k_proj(key_value_states), -1, bsz)
+            value_states = self._shape(self.v_proj(key_value_states), -1, bsz)
+        elif past_key_value is not None:
+            # reuse k, v, self_attention
+            key_states = self._shape(self.k_proj(hidden_states), -1, bsz)
+            value_states = self._shape(self.v_proj(hidden_states), -1, bsz)
+            key_states = torch.cat([past_key_value[0], key_states], dim=2)
+            value_states = torch.cat([past_key_value[1], value_states], dim=2)
+        else:
+            # self_attention
+            key_states = self._shape(self.k_proj(hidden_states), -1, bsz)
+            value_states = self._shape(self.v_proj(hidden_states), -1, bsz)
+
+        if self.is_decoder:
+            # if cross_attention save Tuple(torch.Tensor, torch.Tensor) of all cross attention key/value_states.
+            # Further calls to cross_attention layer can then reuse all cross-attention
+            # key/value_states (first "if" case)
+            # if uni-directional self-attention (decoder) save Tuple(torch.Tensor, torch.Tensor) of
+            # all previous decoder key/value_states. Further calls to uni-directional self-attention
+            # can concat previous decoder key/value_states to current projected key/value_states (third "elif" case)
+            # if encoder bi-directional self-attention `past_key_value` is always `None`
+            past_key_value = (key_states, value_states)
+
+        proj_shape = (bsz * self.num_heads, -1, self.head_dim)
+        query_states = self._shape(query_states, tgt_len, bsz).view(*proj_shape)
+        key_states = key_states.reshape(*proj_shape)
+        value_states = value_states.reshape(*proj_shape)
+
+        src_len = key_states.size(1)
+        attn_weights = torch.bmm(query_states, key_states.transpose(1, 2))
+
+        if attn_weights.size() != (bsz * self.num_heads, tgt_len, src_len):
+            raise ValueError(
+                f"Attention weights should be of size {(bsz * self.num_heads, tgt_len, src_len)}, but is"
+                f" {attn_weights.size()}"
+            )
+
+        if attention_mask is not None:
+            if attention_mask.size() != (bsz, 1, tgt_len, src_len):
+                raise ValueError(
+                    f"Attention mask should be of size {(bsz, 1, tgt_len, src_len)}, but is {attention_mask.size()}"
+                )
+            attn_weights = attn_weights.view(bsz, self.num_heads, tgt_len, src_len) + attention_mask
+            attn_weights = attn_weights.view(bsz * self.num_heads, tgt_len, src_len)
+
+        attn_weights = nn.functional.softmax(attn_weights, dim=-1)
+
+        if layer_head_mask is not None:
+            if layer_head_mask.size() != (self.num_heads,):
+                raise ValueError(
+                    f"Head mask for a single layer should be of size {(self.num_heads,)}, but is"
+                    f" {layer_head_mask.size()}"
+                )
+            attn_weights = layer_head_mask.view(1, -1, 1, 1) * attn_weights.view(bsz, self.num_heads, tgt_len, src_len)
+            attn_weights = attn_weights.view(bsz * self.num_heads, tgt_len, src_len)
+
+        if output_attentions:
+            # this operation is a bit awkward, but it's required to
+            # make sure that attn_weights keeps its gradient.
+            # In order to do so, attn_weights have to be reshaped
+            # twice and have to be reused in the following
+            attn_weights_reshaped = attn_weights.view(bsz, self.num_heads, tgt_len, src_len)
+            attn_weights = attn_weights_reshaped.view(bsz * self.num_heads, tgt_len, src_len)
+        else:
+            attn_weights_reshaped = None
+
+        attn_probs = nn.functional.dropout(attn_weights, p=self.dropout, training=self.training)
+
+        attn_output = torch.bmm(attn_probs, value_states)
+
+        if attn_output.size() != (bsz * self.num_heads, tgt_len, self.head_dim):
+            raise ValueError(
+                f"`attn_output` should be of size {(bsz * self.num_heads, tgt_len, self.head_dim)}, but is"
+                f" {attn_output.size()}"
+            )
+
+        attn_output = attn_output.view(bsz, self.num_heads, tgt_len, self.head_dim)
+        attn_output = attn_output.transpose(1, 2)
+
+        # Use the `embed_dim` from the config (stored in the class) rather than `hidden_state` because `attn_output` can be
+        # partitioned across GPUs when using tensor-parallelism.
+        attn_output = attn_output.reshape(bsz, tgt_len, self.embed_dim)
+
+        attn_output = self.out_proj(attn_output)
+
+        return attn_output, attn_weights_reshaped, past_key_value
+
+
+# Copied from transformers.models.mbart.modeling_mbart.MBartEncoderLayer with MBart->Whisper
+class WhisperEncoderLayer(nn.Module):
+    def __init__(self, config: WhisperConfig):
+        super().__init__()
+        self.embed_dim = config.d_model
+        self.self_attn = WhisperAttention(
+            embed_dim=self.embed_dim,
+            num_heads=config.encoder_attention_heads,
+            dropout=config.attention_dropout,
+        )
+        self.self_attn_layer_norm = nn.LayerNorm(self.embed_dim)
+        self.dropout = config.dropout
+        self.activation_fn = ACT2FN[config.activation_function]
+        self.activation_dropout = config.activation_dropout
+        self.fc1 = nn.Linear(self.embed_dim, config.encoder_ffn_dim)
+        self.fc2 = nn.Linear(config.encoder_ffn_dim, self.embed_dim)
+        self.final_layer_norm = nn.LayerNorm(self.embed_dim)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: torch.Tensor,
+        layer_head_mask: torch.Tensor,
+        output_attentions: bool = False,
+    ) -> torch.Tensor:
+        """
+        Args:
+            hidden_states (`torch.FloatTensor`): input to the layer of shape `(seq_len, batch, embed_dim)`
+            attention_mask (`torch.FloatTensor`): attention mask of size
+                `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
+            layer_head_mask (`torch.FloatTensor`): mask for attention heads in a given layer of size
+                `(encoder_attention_heads,)`.
+            output_attentions (`bool`, *optional*):
+                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
+                returned tensors for more detail.
+        """
+        residual = hidden_states
+        hidden_states = self.self_attn_layer_norm(hidden_states)
+        hidden_states, attn_weights, _ = self.self_attn(
+            hidden_states=hidden_states,
+            attention_mask=attention_mask,
+            layer_head_mask=layer_head_mask,
+            output_attentions=output_attentions,
+        )
+        hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
+        hidden_states = residual + hidden_states
+
+        residual = hidden_states
+        hidden_states = self.final_layer_norm(hidden_states)
+        hidden_states = self.activation_fn(self.fc1(hidden_states))
+        hidden_states = nn.functional.dropout(hidden_states, p=self.activation_dropout, training=self.training)
+        hidden_states = self.fc2(hidden_states)
+        hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
+        hidden_states = residual + hidden_states
+
+        if hidden_states.dtype == torch.float16 and (
+            torch.isinf(hidden_states).any() or torch.isnan(hidden_states).any()
+        ):
+            clamp_value = torch.finfo(hidden_states.dtype).max - 1000
+            hidden_states = torch.clamp(hidden_states, min=-clamp_value, max=clamp_value)
+
+        outputs = (hidden_states,)
+
+        if output_attentions:
+            outputs += (attn_weights,)
+
+        return outputs
+
+
+# Copied from transformers.models.mbart.modeling_mbart.MBartDecoderLayer with MBart->Whisper
+class WhisperDecoderLayer(nn.Module):
+    def __init__(self, config: WhisperConfig):
+        super().__init__()
+        self.embed_dim = config.d_model
+
+        self.self_attn = WhisperAttention(
+            embed_dim=self.embed_dim,
+            num_heads=config.decoder_attention_heads,
+            dropout=config.attention_dropout,
+            is_decoder=True,
+        )
+        self.dropout = config.dropout
+        self.activation_fn = ACT2FN[config.activation_function]
+        self.activation_dropout = config.activation_dropout
+
+        self.self_attn_layer_norm = nn.LayerNorm(self.embed_dim)
+        self.encoder_attn = WhisperAttention(
+            self.embed_dim,
+            config.decoder_attention_heads,
+            dropout=config.attention_dropout,
+            is_decoder=True,
+        )
+        self.encoder_attn_layer_norm = nn.LayerNorm(self.embed_dim)
+        self.fc1 = nn.Linear(self.embed_dim, config.decoder_ffn_dim)
+        self.fc2 = nn.Linear(config.decoder_ffn_dim, self.embed_dim)
+        self.final_layer_norm = nn.LayerNorm(self.embed_dim)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        encoder_hidden_states: Optional[torch.Tensor] = None,
+        encoder_attention_mask: Optional[torch.Tensor] = None,
+        layer_head_mask: Optional[torch.Tensor] = None,
+        cross_attn_layer_head_mask: Optional[torch.Tensor] = None,
+        past_key_value: Optional[Tuple[torch.Tensor]] = None,
+        output_attentions: Optional[bool] = False,
+        use_cache: Optional[bool] = True,
+    ) -> torch.Tensor:
+        """
+        Args:
+            hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
+            attention_mask (`torch.FloatTensor`): attention mask of size
+                `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
+            encoder_hidden_states (`torch.FloatTensor`):
+                cross attention input to the layer of shape `(batch, seq_len, embed_dim)`
+            encoder_attention_mask (`torch.FloatTensor`): encoder attention mask of size
+                `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
+            layer_head_mask (`torch.FloatTensor`): mask for attention heads in a given layer of size
+                `(encoder_attention_heads,)`.
+            cross_attn_layer_head_mask (`torch.FloatTensor`): mask for cross-attention heads in a given layer of
+                size `(decoder_attention_heads,)`.
+            past_key_value (`Tuple(torch.FloatTensor)`): cached past key and value projection states
+            output_attentions (`bool`, *optional*):
+                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
+                returned tensors for more detail.
+        """
+        residual = hidden_states
+        hidden_states = self.self_attn_layer_norm(hidden_states)
+
+        # Self Attention
+        # decoder uni-directional self-attention cached key/values tuple is at positions 1,2
+        self_attn_past_key_value = past_key_value[:2] if past_key_value is not None else None
+        # add present self-attn cache to positions 1,2 of present_key_value tuple
+        hidden_states, self_attn_weights, present_key_value = self.self_attn(
+            hidden_states=hidden_states,
+            past_key_value=self_attn_past_key_value,
+            attention_mask=attention_mask,
+            layer_head_mask=layer_head_mask,
+            output_attentions=output_attentions,
+        )
+        hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
+        hidden_states = residual + hidden_states
+
+        # Cross-Attention Block
+        cross_attn_present_key_value = None
+        cross_attn_weights = None
+        if encoder_hidden_states is not None:
+            residual = hidden_states
+            hidden_states = self.encoder_attn_layer_norm(hidden_states)
+
+            # cross_attn cached key/values tuple is at positions 3,4 of present_key_value tuple
+            cross_attn_past_key_value = past_key_value[-2:] if past_key_value is not None else None
+            hidden_states, cross_attn_weights, cross_attn_present_key_value = self.encoder_attn(
+                hidden_states=hidden_states,
+                key_value_states=encoder_hidden_states,
+                attention_mask=encoder_attention_mask,
+                layer_head_mask=cross_attn_layer_head_mask,
+                past_key_value=cross_attn_past_key_value,
+                output_attentions=output_attentions,
+            )
+            hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
+            hidden_states = residual + hidden_states
+
+            # add cross-attn to positions 3,4 of present_key_value tuple
+            #present_key_value = present_key_value + cross_attn_present_key_value
+
+        # Fully Connected
+        residual = hidden_states
+        hidden_states = self.final_layer_norm(hidden_states)
+        hidden_states = self.activation_fn(self.fc1(hidden_states))
+        hidden_states = nn.functional.dropout(hidden_states, p=self.activation_dropout, training=self.training)
+        hidden_states = self.fc2(hidden_states)
+        hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
+        hidden_states = residual + hidden_states
+
+        outputs = (hidden_states,)
+
+        if output_attentions:
+            outputs += (self_attn_weights, cross_attn_weights)
+
+        if use_cache:
+            outputs += (present_key_value,)
+
+        return outputs
+
+
+class WhisperPreTrainedModel(PreTrainedModel):
+    config_class = WhisperConfig
+    base_model_prefix = "model"
+    main_input_name = "input_features"
+    supports_gradient_checkpointing = True
+    _no_split_modules = ["WhisperEncoderLayer", "WhisperDecoderLayer"]
+
+    def _init_weights(self, module):
+        std = self.config.init_std
+        if isinstance(module, (nn.Linear, nn.Conv1d)):
+            module.weight.data.normal_(mean=0.0, std=std)
+            if module.bias is not None:
+                module.bias.data.zero_()
+        elif isinstance(module, nn.Embedding):
+            module.weight.data.normal_(mean=0.0, std=std)
+            if module.padding_idx is not None:
+                module.weight.data[module.padding_idx].zero_()
+
+    def _set_gradient_checkpointing(self, module, value=False):
+        if isinstance(module, (WhisperDecoder, WhisperEncoder)):
+            module.gradient_checkpointing = value
+
+    def _get_feat_extract_output_lengths(self, input_lengths: torch.LongTensor):
+        """
+        Computes the output length of the convolutional layers
+        """
+        input_lengths = (input_lengths - 1) // 2 + 1
+
+        return input_lengths
+
+
+WHISPER_START_DOCSTRING = r"""
+    This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
+    library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
+    etc.)
+
+    This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
+    Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
+    and behavior.
+
+    Parameters:
+        config ([`WhisperConfig`]):
+            Model configuration class with all the parameters of the model. Initializing with a config file does not
+            load the weights associated with the model, only the configuration. Check out the
+            [`~PreTrainedModel.from_pretrained`] method to load the model weights.
+"""
+
+WHISPER_INPUTS_DOCSTRING = r"""
+    Args:
+        input_features (`torch.FloatTensor` of shape `(batch_size, feature_size, sequence_length)`):
+            Float values mel features extracted from the raw speech waveform. Raw speech waveform can be obtained by
+            loading a `.flac` or `.wav` audio file into an array of type `List[float]` or a `numpy.ndarray`, *e.g.* via
+            the soundfile library (`pip install soundfile`). To prepare the array into `input_features`, the
+            [`AutoFeatureExtractor`] should be used for extracting the mel features, padding and conversion into a
+            tensor of type `torch.FloatTensor`. See [`~WhisperFeatureExtractor.__call__`]
+        attention_mask (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Mask to avoid performing *SpecAugment* data augmentation on padding token indices. Mask values selected in
+            `[0, 1]`:
+
+            - 1 for tokens that are **not masked**,
+            - 0 for tokens that are **masked**.
+
+            [What are attention masks?](../glossary#attention-mask)
+        decoder_input_ids (`torch.LongTensor` of shape `(batch_size, target_sequence_length)`, *optional*):
+            Indices of decoder input sequence tokens in the vocabulary.
+
+            Indices can be obtained using [`WhisperTokenizer`]. See [`PreTrainedTokenizer.encode`] and
+            [`PreTrainedTokenizer.__call__`] for details.
+
+            [What are decoder input IDs?](../glossary#decoder-input-ids)
+
+            Whisper uses the `decoder_start_token_id` as the starting token for `decoder_input_ids` generation. If
+            `past_key_values` is used, optionally only the last `decoder_input_ids` have to be input (see
+            `past_key_values`).
+        decoder_attention_mask (`torch.LongTensor` of shape `(batch_size, target_sequence_length)`, *optional*):
+            Default behavior: generate a tensor that ignores pad tokens in `decoder_input_ids`. Causal mask will also
+            be used by default.
+
+            If you want to change padding behavior, you should read
+            [`modeling_whisper._prepare_decoder_attention_mask`] and modify to your needs. See diagram 1 in [the BART
+            paper](https://arxiv.org/abs/1910.13461) for more information on the default strategy.
+        head_mask (`torch.Tensor` of shape `(encoder_layers, encoder_attention_heads)`, *optional*):
+            Mask to nullify selected heads of the attention modules in the encoder. Mask values selected in `[0, 1]`:
+
+            - 1 indicates the head is **not masked**,
+            - 0 indicates the head is **masked**.
+
+        decoder_head_mask (`torch.Tensor` of shape `(decoder_layers, decoder_attention_heads)`, *optional*):
+            Mask to nullify selected heads of the attention modules in the decoder. Mask values selected in `[0, 1]`:
+
+            - 1 indicates the head is **not masked**,
+            - 0 indicates the head is **masked**.
+
+        cross_attn_head_mask (`torch.Tensor` of shape `(decoder_layers, decoder_attention_heads)`, *optional*):
+            Mask to nullify selected heads of the cross-attention modules. Mask values selected in `[0, 1]`:
+
+            - 1 indicates the head is **not masked**,
+            - 0 indicates the head is **masked**.
+
+        encoder_outputs (`tuple(tuple(torch.FloatTensor)`, *optional*):
+            Tuple consists of (`last_hidden_state`, *optional*: `hidden_states`, *optional*: `attentions`)
+            `last_hidden_state` of shape `(batch_size, sequence_length, hidden_size)`, *optional*) is a sequence of
+            hidden-states at the output of the last layer of the encoder. Used in the cross-attention of the decoder.
+        past_key_values (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
+            Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of shape
+            `(batch_size, num_heads, sequence_length, embed_size_per_head)`) and 2 additional tensors of shape
+            `(batch_size, num_heads, encoder_sequence_length, embed_size_per_head)`.
+
+            Contains pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention
+            blocks) that can be used (see `past_key_values` input) to speed up sequential decoding.
+
+            If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those that
+            don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of all
+            `decoder_input_ids` of shape `(batch_size, sequence_length)`.
+        decoder_inputs_embeds (`torch.FloatTensor` of shape `(batch_size, target_sequence_length, hidden_size)`, *optional*):
+            Optionally, instead of passing `decoder_input_ids` you can choose to directly pass an embedded
+            representation. If `past_key_values` is used, optionally only the last `decoder_inputs_embeds` have to be
+            input (see `past_key_values`). This is useful if you want more control over how to convert
+            `decoder_input_ids` indices into associated vectors than the model's internal embedding lookup matrix.
+        use_cache (`bool`, *optional*):
+            If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see
+            `past_key_values`).
+        output_attentions (`bool`, *optional*):
+            Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
+            tensors for more detail.
+        output_hidden_states (`bool`, *optional*):
+            Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
+            more detail.
+        return_dict (`bool`, *optional*):
+            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+"""
+
+WHISPER_ENCODER_INPUTS_DOCSTRING = r"""
+    Args:
+        input_features (`torch.FloatTensor` of shape `(batch_size, feature_size, sequence_length)`):
+            Float values mel features extracted from the raw speech waveform. Raw speech waveform can be obtained by
+            loading a `.flac` or `.wav` audio file into an array of type `List[float]` or a `numpy.ndarray`, *e.g.* via
+            the soundfile library (`pip install soundfile`). To prepare the array into `input_features`, the
+            [`AutoFeatureExtractor`] should be used for extracting the mel features, padding and conversion into a
+            tensor of type `torch.FloatTensor`. See [`~WhisperFeatureExtractor.__call__`]
+        head_mask (`torch.Tensor` of shape `(encoder_layers, encoder_attention_heads)`, *optional*):
+            Mask to nullify selected heads of the attention modules in the encoder. Mask values selected in `[0, 1]`:
+
+            - 1 indicates the head is **not masked**,
+            - 0 indicates the head is **masked**.
+        encoder_outputs (`tuple(tuple(torch.FloatTensor)`, *optional*):
+            Tuple consists of (`last_hidden_state`, *optional*: `hidden_states`, *optional*: `attentions`)
+            `last_hidden_state` of shape `(batch_size, sequence_length, hidden_size)`, *optional*) is a sequence of
+            hidden-states at the output of the last layer of the encoder.
+        output_attentions (`bool`, *optional*):
+            Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
+            tensors for more detail.
+        output_hidden_states (`bool`, *optional*):
+            Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
+            more detail.
+        return_dict (`bool`, *optional*):
+            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+"""
+
+
+class WhisperEncoder(WhisperPreTrainedModel):
+    """
+    Transformer encoder consisting of *config.encoder_layers* self attention layers. Each layer is a
+    [`WhisperEncoderLayer`].
+
+    Args:
+        config: WhisperConfig
+    """
+
+    def __init__(self, config: WhisperConfig):
+        super().__init__(config)
+        self.dropout = config.dropout
+        self.layerdrop = config.encoder_layerdrop
+
+        embed_dim = config.d_model
+        self.num_mel_bins = config.num_mel_bins
+        self.padding_idx = config.pad_token_id
+        self.max_source_positions = config.max_source_positions
+        self.embed_scale = math.sqrt(embed_dim) if config.scale_embedding else 1.0
+
+        self.conv1 = nn.Conv1d(self.num_mel_bins, embed_dim, kernel_size=3, padding=1)
+        self.conv2 = nn.Conv1d(embed_dim, embed_dim, kernel_size=3, stride=2, padding=1)
+
+        self.embed_positions = nn.Embedding(self.max_source_positions, embed_dim)
+
+        self.layers = nn.ModuleList([WhisperEncoderLayer(config) for _ in range(config.encoder_layers)])
+        self.layer_norm = nn.LayerNorm(config.d_model)
+
+        self.gradient_checkpointing = False
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def _freeze_parameters(self):
+        for param in self.parameters():
+            param.requires_grad = False
+        self._requires_grad = False
+
+    def get_input_embeddings(self) -> nn.Module:
+        return self.conv1
+
+    def set_input_embeddings(self, value: nn.Module):
+        self.conv1 = value
+
+    def forward(
+        self,
+        input_features,
+        attention_mask=None,
+        head_mask=None,
+        output_attentions=None,
+        output_hidden_states=None,
+        return_dict=None,
+    ):
+        r"""
+        Args:
+            input_features (`torch.LongTensor` of shape `(batch_size, feature_size, sequence_length)`):
+                Float values of mel features extracted from the raw speech waveform. Raw speech waveform can be
+                obtained by loading a `.flac` or `.wav` audio file into an array of type `List[float]` or a
+                `numpy.ndarray`, *e.g.* via the soundfile library (`pip install soundfile`). To prepare the array into
+                `input_features`, the [`AutoFeatureExtractor`] should be used for extracting the mel features, padding
+                and conversion into a tensor of type `torch.FloatTensor`. See [`~WhisperFeatureExtractor.__call__`]
+            attention_mask (`torch.Tensor`)`, *optional*):
+                Whisper does not support masking of the `input_features`, this argument is preserved for compatibility,
+                but it is not used. By default the silence in the input log mel spectrogram are ignored.
+            head_mask (`torch.Tensor` of shape `(encoder_layers, encoder_attention_heads)`, *optional*):
+                Mask to nullify selected heads of the attention modules. Mask values selected in `[0, 1]`:
+
+                - 1 indicates the head is **not masked**,
+                - 0 indicates the head is **masked**.
+            output_attentions (`bool`, *optional*):
+                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
+                returned tensors for more detail.
+            output_hidden_states (`bool`, *optional*):
+                Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors
+                for more detail.
+            return_dict (`bool`, *optional*):
+                Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+        """
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+        inputs_embeds = nn.functional.gelu(self.conv1(input_features))
+        inputs_embeds = nn.functional.gelu(self.conv2(inputs_embeds))
+
+        inputs_embeds = inputs_embeds.permute(0, 2, 1)
+        embed_pos = self.embed_positions.weight
+
+        hidden_states = inputs_embeds + embed_pos
+        hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
+
+        encoder_states = () if output_hidden_states else None
+        all_attentions = () if output_attentions else None
+
+        # check if head_mask has a correct number of layers specified if desired
+        if head_mask is not None:
+            assert head_mask.size()[0] == (
+                len(self.layers)
+            ), f"The head_mask should be specified for {len(self.layers)} layers, but it is for {head_mask.size()[0]}."
+
+        for idx, encoder_layer in enumerate(self.layers):
+            if output_hidden_states:
+                encoder_states = encoder_states + (hidden_states,)
+            # add LayerDrop (see https://arxiv.org/abs/1909.11556 for description)
+            dropout_probability = random.uniform(0, 1)
+            if self.training and (dropout_probability < self.layerdrop):  # skip the layer
+                layer_outputs = (None, None)
+            else:
+                if self.gradient_checkpointing and self.training:
+
+                    def create_custom_forward(module):
+                        def custom_forward(*inputs):
+                            return module(*inputs, output_attentions)
+
+                        return custom_forward
+
+                    layer_outputs = torch.utils.checkpoint.checkpoint(
+                        create_custom_forward(encoder_layer),
+                        hidden_states,
+                        None,
+                        (head_mask[idx] if head_mask is not None else None),
+                    )
+                else:
+                    layer_outputs = encoder_layer(
+                        hidden_states,
+                        None,
+                        layer_head_mask=(head_mask[idx] if head_mask is not None else None),
+                        output_attentions=output_attentions,
+                    )
+
+                hidden_states = layer_outputs[0]
+
+            if output_attentions:
+                all_attentions = all_attentions + (layer_outputs[1],)
+
+        hidden_states = self.layer_norm(hidden_states)
+        if output_hidden_states:
+            encoder_states = encoder_states + (hidden_states,)
+
+        if not return_dict:
+            return tuple(v for v in [hidden_states, encoder_states, all_attentions] if v is not None)
+        return BaseModelOutput(
+            last_hidden_state=hidden_states, hidden_states=encoder_states, attentions=all_attentions
+        )
+
+
+class WhisperDecoder(WhisperPreTrainedModel):
+    """
+    Transformer decoder consisting of *config.decoder_layers* layers. Each layer is a [`WhisperDecoderLayer`]
+
+    Args:
+        config: WhisperConfig
+    """
+
+    def __init__(self, config: WhisperConfig):
+        super().__init__(config)
+        self.dropout = config.dropout
+        self.layerdrop = config.decoder_layerdrop
+        self.padding_idx = config.pad_token_id
+        self.max_target_positions = config.max_target_positions
+        self.max_source_positions = config.max_source_positions
+        self.embed_scale = math.sqrt(config.d_model) if config.scale_embedding else 1.0
+
+        self.embed_tokens = nn.Embedding(config.vocab_size, config.d_model, self.padding_idx)
+        self.embed_positions = WhisperPositionalEmbedding(self.max_target_positions, config.d_model)
+
+        self.layers = nn.ModuleList([WhisperDecoderLayer(config) for _ in range(config.decoder_layers)])
+
+        self.layer_norm = nn.LayerNorm(config.d_model)
+
+        self.gradient_checkpointing = False
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.embed_tokens = value
+
+    def _prepare_decoder_attention_mask(self, attention_mask, input_shape, inputs_embeds, past_key_values_length):
+        # create causal mask
+        # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
+        combined_attention_mask = None
+
+        if input_shape[-1] > 1:
+            combined_attention_mask = _make_causal_mask(
+                input_shape,
+                inputs_embeds.dtype,
+                device=inputs_embeds.device,
+                past_key_values_length=past_key_values_length,
+            )
+
+        if attention_mask is not None:
+            # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
+            expanded_attn_mask = _expand_mask(attention_mask, inputs_embeds.dtype, tgt_len=input_shape[-1])
+            combined_attention_mask = (
+                expanded_attn_mask if combined_attention_mask is None else expanded_attn_mask + combined_attention_mask
+            )
+
+        return combined_attention_mask
+
+    def forward(
+        self,
+        input_ids=None,
+        attention_mask=None,
+        encoder_hidden_states=None,
+        head_mask=None,
+        cross_attn_head_mask=None,
+        past_key_values=None,
+        inputs_embeds=None,
+        use_cache=None,
+        output_attentions=None,
+        output_hidden_states=None,
+        return_dict=None,
+    ):
+        r"""
+        Args:
+            input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
+                Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you
+                provide it.
+
+                Indices can be obtained using [`WhisperTokenizer`]. See [`PreTrainedTokenizer.encode`] and
+                [`PreTrainedTokenizer.__call__`] for details.
+
+                [What are input IDs?](../glossary#input-ids)
+            attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
+                Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
+
+                - 1 for tokens that are **not masked**,
+                - 0 for tokens that are **masked**.
+
+                [What are attention masks?](../glossary#attention-mask)
+            encoder_hidden_states (`torch.FloatTensor` of shape `(batch_size, encoder_sequence_length, hidden_size)`, *optional*):
+                Sequence of hidden-states at the output of the last layer of the encoder. Used in the cross-attention
+                of the decoder.
+            head_mask (`torch.Tensor` of shape `(decoder_layers, decoder_attention_heads)`, *optional*):
+                Mask to nullify selected heads of the attention modules. Mask values selected in `[0, 1]`:
+
+                - 1 indicates the head is **not masked**,
+                - 0 indicates the head is **masked**.
+
+            cross_attn_head_mask (`torch.Tensor` of shape `(decoder_layers, decoder_attention_heads)`, *optional*):
+                Mask to nullify selected heads of the attention modules in encoder to avoid performing cross-attention
+                on hidden heads. Mask values selected in `[0, 1]`:
+
+                - 1 indicates the head is **not masked**,
+                - 0 indicates the head is **masked**.
+
+            past_key_values (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
+                Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of
+                shape `(batch_size, num_heads, sequence_length, embed_size_per_head)`) and 2 additional tensors of
+                shape `(batch_size, num_heads, encoder_sequence_length, embed_size_per_head)`.
+
+                Contains pre-computed hidden-states (key and values in the self-attention blocks and in the
+                cross-attention blocks) that can be used (see `past_key_values` input) to speed up sequential decoding.
+
+                If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those
+                that don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of
+                all `decoder_input_ids` of shape `(batch_size, sequence_length)`. inputs_embeds (`torch.FloatTensor` of
+                shape `(batch_size, sequence_length, hidden_size)`, *optional*): Optionally, instead of passing
+                `input_ids` you can choose to directly pass an embedded representation. This is useful if you want more
+                control over how to convert `input_ids` indices into associated vectors than the model's internal
+                embedding lookup matrix.
+            output_attentions (`bool`, *optional*):
+                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
+                returned tensors for more detail.
+            output_hidden_states (`bool`, *optional*):
+                Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors
+                for more detail.
+            return_dict (`bool`, *optional*):
+                Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+        """
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        use_cache = use_cache if use_cache is not None else self.config.use_cache
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        # retrieve input_ids and inputs_embeds
+        if input_ids is not None and inputs_embeds is not None:
+            raise ValueError("You cannot specify both decoder_input_ids and decoder_inputs_embeds at the same time")
+        elif input_ids is not None:
+            input_shape = input_ids.size()
+            input_ids = input_ids.view(-1, input_shape[-1])
+        elif inputs_embeds is not None:
+            input_shape = inputs_embeds.size()[:-1]
+        else:
+            raise ValueError("You have to specify either decoder_input_ids or decoder_inputs_embeds")
+
+        # past_key_values_length
+        # MODIFIED: handle BOTH cache layouts — the MLA latent cache is 3-dim (bsz, seq, latent)
+        # -> seq at dim 1; the stock MHA cache is 4-dim (bsz, heads, seq, head_dim) -> seq at
+        # dim 2 (upstream hardcoded shape[1], which breaks stock incremental decoding).
+        if past_key_values is not None:
+            _first = past_key_values[0][0]
+            past_key_values_length = _first.shape[1] if _first.dim() == 3 else _first.shape[2]
+        else:
+            past_key_values_length = 0
+
+        if inputs_embeds is None:
+            inputs_embeds = self.embed_tokens(input_ids)
+
+        attention_mask = self._prepare_decoder_attention_mask(
+            attention_mask, input_shape, inputs_embeds, past_key_values_length
+        )
+
+        # embed positions
+        if input_ids is not None:
+            positions = self.embed_positions(input_ids, past_key_values_length=past_key_values_length)
+        else:
+            positions = self.embed_positions(inputs_embeds, past_key_values_length=past_key_values_length)
+
+        hidden_states = inputs_embeds + positions
+        hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
+
+        if self.gradient_checkpointing and self.training:
+            if use_cache:
+                logger.warning_once(
+                    "`use_cache = True` is incompatible with gradient checkpointing. Setting `use_cache = False`..."
+                )
+                use_cache = False
+        # decoder layers
+        all_hidden_states = () if output_hidden_states else None
+        all_self_attns = () if output_attentions else None
+        all_cross_attentions = () if (output_attentions and encoder_hidden_states is not None) else None
+        next_decoder_cache = () if use_cache else None
+
+        # check if head_mask/cross_attn_head_mask has a correct number of layers specified if desired
+        for attn_mask, mask_name in zip([head_mask, cross_attn_head_mask], ["head_mask", "cross_attn_head_mask"]):
+            if attn_mask is not None:
+                assert attn_mask.size()[0] == (len(self.layers)), (
+                    f"The `{mask_name}` should be specified for {len(self.layers)} layers, but it is for"
+                    f" {head_mask.size()[0]}."
+                )
+        for idx, decoder_layer in enumerate(self.layers):
+            # add LayerDrop (see https://arxiv.org/abs/1909.11556 for description)
+            if output_hidden_states:
+                all_hidden_states += (hidden_states,)
+            dropout_probability = random.uniform(0, 1)
+            if self.training and (dropout_probability < self.layerdrop):
+                continue
+
+            past_key_value = past_key_values[idx] if past_key_values is not None else None
+
+            if self.gradient_checkpointing and self.training:
+
+                def create_custom_forward(module):
+                    def custom_forward(*inputs):
+                        # None for past_key_value
+                        return module(*inputs, output_attentions, use_cache)
+
+                    return custom_forward
+
+                layer_outputs = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(decoder_layer),
+                    hidden_states,
+                    attention_mask,
+                    encoder_hidden_states,
+                    None,  # encoder attention mask
+                    head_mask[idx] if head_mask is not None else None,
+                    cross_attn_head_mask[idx] if cross_attn_head_mask is not None else None,
+                    None,  # past_key_value
+                )
+            else:
+                layer_outputs = decoder_layer(
+                    hidden_states,
+                    attention_mask=attention_mask,
+                    encoder_hidden_states=encoder_hidden_states,
+                    layer_head_mask=(head_mask[idx] if head_mask is not None else None),
+                    cross_attn_layer_head_mask=(
+                        cross_attn_head_mask[idx] if cross_attn_head_mask is not None else None
+                    ),
+                    past_key_value=past_key_value,
+                    output_attentions=output_attentions,
+                    use_cache=use_cache,
+                )
+            hidden_states = layer_outputs[0]
+
+            if use_cache:
+                next_decoder_cache += (layer_outputs[3 if output_attentions else 1],)
+
+            if output_attentions:
+                all_self_attns += (layer_outputs[1],)
+
+                if encoder_hidden_states is not None:
+                    all_cross_attentions += (layer_outputs[2],)
+
+        hidden_states = self.layer_norm(hidden_states)
+        # add hidden states from the last decoder layer
+        if output_hidden_states:
+            all_hidden_states += (hidden_states,)
+
+        next_cache = next_decoder_cache if use_cache else None
+        if not return_dict:
+            return tuple(
+                v
+                for v in [hidden_states, next_cache, all_hidden_states, all_self_attns, all_cross_attentions]
+                if v is not None
+            )
+        return BaseModelOutputWithPastAndCrossAttentions(
+            last_hidden_state=hidden_states,
+            past_key_values=next_cache,
+            hidden_states=all_hidden_states,
+            attentions=all_self_attns,
+            cross_attentions=all_cross_attentions,
+        )
+
+
+@add_start_docstrings(
+    "The bare Whisper Model outputting raw hidden-states without any specific head on top.",
+    WHISPER_START_DOCSTRING,
+)
+class WhisperModel(WhisperPreTrainedModel):
+    _keys_to_ignore_on_load_missing = [r"proj_out.weight"]
+
+    def __init__(self, config: WhisperConfig):
+        super().__init__(config)
+
+        self.encoder = WhisperEncoder(config)
+        self.decoder = WhisperDecoder(config)
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.decoder.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.decoder.embed_tokens = value
+
+    def get_encoder(self):
+        return self.encoder
+
+    def get_decoder(self):
+        return self.decoder
+
+    def freeze_encoder(self):
+        """
+        Calling this function will disable the gradient computation for the Whisper encoder so that its parameters will
+        not be updated during training.
+        """
+        self.encoder._freeze_parameters()
+
+    def _mask_input_features(
+        self,
+        input_features: torch.FloatTensor,
+        attention_mask: Optional[torch.LongTensor] = None,
+    ):
+        """
+        Masks extracted features along time axis and/or along feature axis according to
+        [SpecAugment](https://arxiv.org/abs/1904.08779).
+        """
+
+        # `config.apply_spec_augment` can set masking to False
+        if not getattr(self.config, "apply_spec_augment", True):
+            return input_features
+
+        # generate indices & apply SpecAugment along time axis
+        batch_size, hidden_size, sequence_length = input_features.size()
+
+        if self.config.mask_time_prob > 0 and self.training:
+            # generate indices & apply SpecAugment along time axis
+            mask_time_indices = _compute_mask_indices(
+                (batch_size, sequence_length),
+                mask_prob=self.config.mask_time_prob,
+                mask_length=self.config.mask_time_length,
+                attention_mask=attention_mask,
+                min_masks=self.config.mask_time_min_masks,
+            )
+            mask_time_indices = torch.tensor(mask_time_indices, device=input_features.device, dtype=torch.bool)
+            mask_time_indices = mask_time_indices[:, None].expand(-1, hidden_size, -1)
+            input_features[mask_time_indices] = 0
+
+        if self.config.mask_feature_prob > 0 and self.training:
+            # generate indices & apply SpecAugment along feature axis
+            mask_feature_indices = _compute_mask_indices(
+                (batch_size, hidden_size),
+                mask_prob=self.config.mask_feature_prob,
+                mask_length=self.config.mask_feature_length,
+                min_masks=self.config.mask_feature_min_masks,
+            )
+            mask_feature_indices = torch.tensor(mask_feature_indices, device=input_features.device, dtype=torch.bool)
+            input_features[mask_feature_indices] = 0
+
+        return input_features
+
+    @add_start_docstrings_to_model_forward(WHISPER_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=Seq2SeqModelOutput, config_class=_CONFIG_FOR_DOC)
+    def forward(
+        self,
+        input_features: Optional[torch.FloatTensor] = None,
+        attention_mask: Optional[torch.LongTensor] = None,
+        decoder_input_ids: Optional[torch.LongTensor] = None,
+        decoder_attention_mask: Optional[torch.LongTensor] = None,
+        head_mask: Optional[torch.Tensor] = None,
+        decoder_head_mask: Optional[torch.Tensor] = None,
+        cross_attn_head_mask: Optional[torch.Tensor] = None,
+        encoder_outputs: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,
+        past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,
+        decoder_inputs_embeds: Optional[Tuple[torch.FloatTensor]] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple[torch.Tensor], Seq2SeqModelOutput]:
+        r"""
+        Returns:
+
+        Example:
+         ```python
+         >>> import torch
+         >>> from transformers import AutoFeatureExtractor, WhisperModel
+         >>> from datasets import load_dataset
+
+         >>> model = WhisperModel.from_pretrained("openai/whisper-base")
+         >>> feature_extractor = AutoFeatureExtractor.from_pretrained("openai/whisper-base")
+         >>> ds = load_dataset("hf-internal-testing/librispeech_asr_dummy", "clean", split="validation")
+         >>> inputs = feature_extractor(ds[0]["audio"]["array"], return_tensors="pt")
+         >>> input_features = inputs.input_features
+         >>> decoder_input_ids = torch.tensor([[1, 1]]) * model.config.decoder_start_token_id
+         >>> last_hidden_state = model(input_features, decoder_input_ids=decoder_input_ids).last_hidden_state
+         >>> list(last_hidden_state.shape)
+         [1, 2, 512]
+         ```"""
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        use_cache = use_cache if use_cache is not None else self.config.use_cache
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        if encoder_outputs is None:
+            input_features = self._mask_input_features(input_features, attention_mask=attention_mask)
+
+            encoder_outputs = self.encoder(
+                input_features,
+                head_mask=head_mask,
+                output_attentions=output_attentions,
+                output_hidden_states=output_hidden_states,
+                return_dict=return_dict,
+            )
+        # If the user passed a tuple for encoder_outputs, we wrap it in a BaseModelOutput when return_dict=True
+        elif return_dict and not isinstance(encoder_outputs, BaseModelOutput):
+            encoder_outputs = BaseModelOutput(
+                last_hidden_state=encoder_outputs[0],
+                hidden_states=encoder_outputs[1] if len(encoder_outputs) > 1 else None,
+                attentions=encoder_outputs[2] if len(encoder_outputs) > 2 else None,
+            )
+
+        # decoder outputs consists of (dec_features, past_key_value, dec_hidden, dec_attn)
+        decoder_outputs = self.decoder(
+            input_ids=decoder_input_ids,
+            attention_mask=decoder_attention_mask,
+            encoder_hidden_states=encoder_outputs[0],
+            head_mask=decoder_head_mask,
+            cross_attn_head_mask=cross_attn_head_mask,
+            past_key_values=past_key_values,
+            inputs_embeds=decoder_inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        if not return_dict:
+            return decoder_outputs + encoder_outputs
+
+        return Seq2SeqModelOutput(
+            last_hidden_state=decoder_outputs.last_hidden_state,
+            past_key_values=decoder_outputs.past_key_values,
+            decoder_hidden_states=decoder_outputs.hidden_states,
+            decoder_attentions=decoder_outputs.attentions,
+            cross_attentions=decoder_outputs.cross_attentions,
+            encoder_last_hidden_state=encoder_outputs.last_hidden_state,
+            encoder_hidden_states=encoder_outputs.hidden_states,
+            encoder_attentions=encoder_outputs.attentions,
+        )
+
+
+@add_start_docstrings(
+    "The Whisper Model with a language modeling head. Can be used for automatic speech recognition.",
+    WHISPER_START_DOCSTRING,
+)
+class WhisperForConditionalGeneration(WhisperPreTrainedModel):
+    base_model_prefix = "model"
+    _keys_to_ignore_on_load_missing = [
+        r"encoder.version",
+        r"decoder.version",
+        r"proj_out.weight",
+    ]
+    _keys_to_ignore_on_save = [
+        r"proj_out.weight",
+    ]
+
+    def __init__(self, config: WhisperConfig):
+        super().__init__(config)
+        self.model = WhisperModel(config)
+        self.proj_out = nn.Linear(config.d_model, config.vocab_size, bias=False)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_encoder(self):
+        return self.model.get_encoder()
+
+    def get_decoder(self):
+        return self.model.get_decoder()
+
+    def resize_token_embeddings(self, new_num_tokens: int) -> nn.Embedding:
+        new_embeddings = super().resize_token_embeddings(new_num_tokens)
+        return new_embeddings
+
+    def get_output_embeddings(self):
+        return self.proj_out
+
+    def set_output_embeddings(self, new_embeddings):
+        self.proj_out = new_embeddings
+
+    def get_input_embeddings(self) -> nn.Module:
+        return self.model.get_input_embeddings()
+
+    def freeze_encoder(self):
+        """
+        Calling this function will disable the gradient computation for the Whisper encoder so that its parameters will
+        not be updated during training.
+        """
+        self.model.encoder._freeze_parameters()
+
+    @add_start_docstrings_to_model_forward(WHISPER_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=Seq2SeqLMOutput, config_class=_CONFIG_FOR_DOC)
+    def forward(
+        self,
+        input_features: Optional[torch.FloatTensor] = None,
+        attention_mask: Optional[torch.LongTensor] = None,
+        decoder_input_ids: Optional[torch.LongTensor] = None,
+        decoder_attention_mask: Optional[torch.LongTensor] = None,
+        head_mask: Optional[torch.Tensor] = None,
+        decoder_head_mask: Optional[torch.Tensor] = None,
+        cross_attn_head_mask: Optional[torch.Tensor] = None,
+        encoder_outputs: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,
+        past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,
+        decoder_inputs_embeds: Optional[Tuple[torch.FloatTensor]] = None,
+        labels: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple[torch.Tensor], Seq2SeqLMOutput]:
+        r"""
+        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Labels for computing the language modeling loss. Indices should either be in `[0, ..., config.vocab_size]`
+            or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored (masked), the loss is
+            only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
+
+        Returns:
+
+        Example:
+
+        ```python
+        >>> import torch
+        >>> from transformers import AutoProcessor, WhisperForConditionalGeneration
+        >>> from datasets import load_dataset
+
+        >>> processor = AutoProcessor.from_pretrained("openai/whisper-tiny.en")
+        >>> model = WhisperForConditionalGeneration.from_pretrained("openai/whisper-tiny.en")
+
+        >>> ds = load_dataset("hf-internal-testing/librispeech_asr_dummy", "clean", split="validation")
+
+        >>> inputs = processor(ds[0]["audio"]["array"], return_tensors="pt")
+        >>> input_features = inputs.input_features
+
+        >>> generated_ids = model.generate(inputs=input_features)
+
+        >>> transcription = processor.batch_decode(generated_ids, skip_special_tokens=True)[0]
+        >>> transcription
+        ' Mr. Quilter is the apostle of the middle classes, and we are glad to welcome his gospel.'
+        ```"""
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        if labels is not None:
+            if decoder_input_ids is None and decoder_inputs_embeds is None:
+                decoder_input_ids = shift_tokens_right(
+                    labels, self.config.pad_token_id, self.config.decoder_start_token_id
+                )
+
+        outputs = self.model(
+            input_features,
+            attention_mask=attention_mask,
+            decoder_input_ids=decoder_input_ids,
+            encoder_outputs=encoder_outputs,
+            decoder_attention_mask=decoder_attention_mask,
+            head_mask=head_mask,
+            decoder_head_mask=decoder_head_mask,
+            cross_attn_head_mask=cross_attn_head_mask,
+            past_key_values=past_key_values,
+            decoder_inputs_embeds=decoder_inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+        lm_logits = self.proj_out(outputs[0])
+
+        loss = None
+        if labels is not None:
+            loss_fct = CrossEntropyLoss()
+            # move labels to correct device to enable PP
+            labels = labels.to(lm_logits.device)
+            loss = loss_fct(lm_logits.view(-1, self.config.vocab_size), labels.reshape(-1))
+
+        if not return_dict:
+            output = (lm_logits,) + outputs[1:]
+            return ((loss,) + output) if loss is not None else output
+
+        return Seq2SeqLMOutput(
+            loss=loss,
+            logits=lm_logits,
+            past_key_values=outputs.past_key_values,
+            decoder_hidden_states=outputs.decoder_hidden_states,
+            decoder_attentions=outputs.decoder_attentions,
+            cross_attentions=outputs.cross_attentions,
+            encoder_last_hidden_state=outputs.encoder_last_hidden_state,
+            encoder_hidden_states=outputs.encoder_hidden_states,
+            encoder_attentions=outputs.encoder_attentions,
+        )
+
+    def generate(
+        self,
+        inputs: Optional[torch.Tensor] = None,
+        generation_config=None,
+        logits_processor=None,
+        stopping_criteria=None,
+        prefix_allowed_tokens_fn=None,
+        synced_gpus=False,
+        return_timestamps=None,
+        task=None,
+        language=None,
+        is_multilingual=None,
+        **kwargs,
+    ):
+        """
+
+        Generates sequences of token ids for models with a language modeling head.
+
+        <Tip warning={true}>
+
+        Most generation-controlling parameters are set in `generation_config` which, if not passed, will be set to the
+        model's default generation configuration. You can override any `generation_config` by passing the corresponding
+        parameters to generate(), e.g. `.generate(inputs, num_beams=4, do_sample=True)`.
+
+        For an overview of generation strategies and code examples, check out the [following
+        guide](./generation_strategies).
+
+        </Tip>
+
+        Parameters:
+            inputs (`torch.Tensor` of varying shape depending on the modality, *optional*):
+                The sequence used as a prompt for the generation or as model inputs to the encoder. If `None` the
+                method initializes it with `bos_token_id` and a batch size of 1. For decoder-only models `inputs`
+                should of in the format of `input_ids`. For encoder-decoder models *inputs* can represent any of
+                `input_ids`, `input_values`, `input_features`, or `pixel_values`.
+            generation_config (`~generation.GenerationConfig`, *optional*):
+                The generation configuration to be used as base parametrization for the generation call. `**kwargs`
+                passed to generate matching the attributes of `generation_config` will override them. If
+                `generation_config` is not provided, the default will be used, which had the following loading
+                priority: 1) from the `generation_config.json` model file, if it exists; 2) from the model
+                configuration. Please note that unspecified parameters will inherit [`~generation.GenerationConfig`]'s
+                default values, whose documentation should be checked to parameterize generation.
+            logits_processor (`LogitsProcessorList`, *optional*):
+                Custom logits processors that complement the default logits processors built from arguments and
+                generation config. If a logit processor is passed that is already created with the arguments or a
+                generation config an error is thrown. This feature is intended for advanced users.
+            stopping_criteria (`StoppingCriteriaList`, *optional*):
+                Custom stopping criteria that complement the default stopping criteria built from arguments and a
+                generation config. If a stopping criteria is passed that is already created with the arguments or a
+                generation config an error is thrown. This feature is intended for advanced users.
+            prefix_allowed_tokens_fn (`Callable[[int, torch.Tensor], List[int]]`, *optional*):
+                If provided, this function constraints the beam search to allowed tokens only at each step. If not
+                provided no constraint is applied. This function takes 2 arguments: the batch ID `batch_id` and
+                `input_ids`. It has to return a list with the allowed tokens for the next generation step conditioned
+                on the batch ID `batch_id` and the previously generated tokens `inputs_ids`. This argument is useful
+                for constrained generation conditioned on the prefix, as described in [Autoregressive Entity
+                Retrieval](https://arxiv.org/abs/2010.00904).
+            synced_gpus (`bool`, *optional*, defaults to `False`):
+                Whether to continue running the while loop until max_length (needed for ZeRO stage 3)
+            return_timestamps (`bool`, *optional*):
+                Whether to return the timestamps with the text. This enables the `WhisperTimestampsLogitsProcessor`.
+            task (`bool`, *optional*):
+                Task to use for generation, either "translate" or "transcribe". The `model.config.forced_decoder_ids`
+                will be updated accordingly.
+            language (`bool`, *optional*):
+                Language token to use for generation, can be either in the form of `<|en|>`, `en` or `english`. You can
+                find all the possible language tokens in the `model.generation_config.lang_to_id` dictionary.
+            is_multilingual (`bool`, *optional*):
+                Whether or not the model is multilingual.
+            kwargs:
+                Ad hoc parametrization of `generate_config` and/or additional model-specific kwargs that will be
+                forwarded to the `forward` function of the model. If the model is an encoder-decoder model, encoder
+                specific kwargs should not be prefixed and decoder specific kwargs should be prefixed with *decoder_*.
+
+        Return:
+            [`~utils.ModelOutput`] or `torch.LongTensor`: A [`~utils.ModelOutput`] (if `return_dict_in_generate=True`
+            or when `config.return_dict_in_generate=True`) or a `torch.FloatTensor`.
+
+                If the model is *not* an encoder-decoder model (`model.config.is_encoder_decoder=False`), the possible
+                [`~utils.ModelOutput`] types are:
+
+                    - [`~generation.GreedySearchDecoderOnlyOutput`],
+                    - [`~generation.SampleDecoderOnlyOutput`],
+                    - [`~generation.BeamSearchDecoderOnlyOutput`],
+                    - [`~generation.BeamSampleDecoderOnlyOutput`]
+
+                If the model is an encoder-decoder model (`model.config.is_encoder_decoder=True`), the possible
+                [`~utils.ModelOutput`] types are:
+
+                    - [`~generation.GreedySearchEncoderDecoderOutput`],
+                    - [`~generation.SampleEncoderDecoderOutput`],
+                    - [`~generation.BeamSearchEncoderDecoderOutput`],
+                    - [`~generation.BeamSampleEncoderDecoderOutput`]
+        """
+        if generation_config is None:
+            generation_config = self.generation_config
+
+        if return_timestamps is not None:
+            if not hasattr(generation_config, "no_timestamps_token_id"):
+                raise ValueError(
+                    "You are trying to return timestamps, but the generation config is not properly set."
+                    "Make sure to initialize the generation config with the correct attributes that are needed such as `no_timestamps_token_id`."
+                    "For more details on how to generate the approtiate config, refer to https://github.com/huggingface/transformers/issues/21878#issuecomment-1451902363"
+                )
+
+            generation_config.return_timestamps = return_timestamps
+        else:
+            generation_config.return_timestamps = False
+
+        if language is not None:
+            language = language.lower()
+            generation_config.language = language
+        if task is not None:
+            generation_config.task = task
+
+        forced_decoder_ids = []
+        if task is not None or language is not None:
+            if hasattr(generation_config, "language"):
+                if generation_config.language in generation_config.lang_to_id.keys():
+                    language_token = generation_config.language
+                elif generation_config.language in TO_LANGUAGE_CODE.keys():
+                    language_token = f"<|{TO_LANGUAGE_CODE[generation_config.language]}|>"
+                elif generation_config.language in TO_LANGUAGE_CODE.values():
+                    language_token = f"<|{generation_config.language}|>"
+                else:
+                    is_language_code = len(generation_config.language) == 2
+                    raise ValueError(
+                        f"Unsupported language: {generation_config.language}. Language should be one of:"
+                        f" {list(TO_LANGUAGE_CODE.values()) if is_language_code else list(TO_LANGUAGE_CODE.keys())}."
+                    )
+                forced_decoder_ids.append((1, generation_config.lang_to_id[language_token]))
+            else:
+                forced_decoder_ids.append((1, None))  # automatically detect the language
+
+            if hasattr(generation_config, "task"):
+                if generation_config.task in TASK_IDS:
+                    forced_decoder_ids.append((2, generation_config.task_to_id[generation_config.task]))
+                else:
+                    raise ValueError(
+                        f"The `{generation_config.task}`task is not supported. The task should be one of `{TASK_IDS}`"
+                    )
+            else:
+                forced_decoder_ids.append((2, generation_config.task_to_id["transcribe"]))  # defaults to transcribe
+            if hasattr(generation_config, "no_timestamps_token_id") and not generation_config.return_timestamps:
+                idx = forced_decoder_ids[-1][0] + 1 if forced_decoder_ids else 1
+                forced_decoder_ids.append((idx, generation_config.no_timestamps_token_id))
+
+        # Legacy code for backward compatibility
+        elif hasattr(self.config, "forced_decoder_ids") and self.config.forced_decoder_ids is not None:
+            forced_decoder_ids = self.config.forced_decoder_ids
+        elif (
+            hasattr(self.generation_config, "forced_decoder_ids")
+            and self.generation_config.forced_decoder_ids is not None
+        ):
+            forced_decoder_ids = self.generation_config.forced_decoder_ids
+
+        if generation_config.return_timestamps:
+            logits_processor = [WhisperTimeStampLogitsProcessor(generation_config)]
+
+        if len(forced_decoder_ids) > 0:
+            generation_config.forced_decoder_ids = forced_decoder_ids
+
+        return super().generate(
+            inputs,
+            generation_config,
+            logits_processor,
+            stopping_criteria,
+            prefix_allowed_tokens_fn,
+            synced_gpus,
+            **kwargs,
+        )
+
+    def prepare_inputs_for_generation(
+        self,
+        decoder_input_ids,
+        past_key_values=None,
+        use_cache=None,
+        encoder_outputs=None,
+        attention_mask=None,
+        **kwargs,
+    ):
+        # cut decoder_input_ids if past is used
+        if past_key_values is not None:
+            decoder_input_ids = decoder_input_ids[:, -1:]
+
+        return {
+            "encoder_outputs": encoder_outputs,
+            "past_key_values": past_key_values,
+            "decoder_input_ids": decoder_input_ids,
+            "use_cache": use_cache,
+            "decoder_attention_mask": None,
+        }
+
+    #
+    @staticmethod
+    def _reorder_cache(past_key_values, beam_idx):
+        reordered_past = ()
+        for layer_past in past_key_values:
+            reordered_past += (tuple(past_state.index_select(0, beam_idx) for past_state in layer_past),)
+        return reordered_past
+
+
+@add_start_docstrings(
+    """
+    Whisper Encoder Model with a sequence classification head on top (a linear layer over the pooled output) for tasks
+    like SUPERB Keyword Spotting.
+    """,
+    WHISPER_ENCODER_INPUTS_DOCSTRING,
+)
+class WhisperForAudioClassification(WhisperPreTrainedModel):
+    def __init__(self, config):
+        super().__init__(config)
+
+        self.encoder = WhisperEncoder(config)
+        num_layers = config.num_hidden_layers + 1  # transformer layers + input embeddings
+        if config.use_weighted_layer_sum:
+            self.layer_weights = nn.Parameter(torch.ones(num_layers) / num_layers)
+        self.projector = nn.Linear(config.hidden_size, config.classifier_proj_size)
+        self.classifier = nn.Linear(config.classifier_proj_size, config.num_labels)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def freeze_encoder(self):
+        """
+        Calling this function will disable the gradient computation for the Whisper encoder so that its parameters will
+        not be updated during training. Only the projection layers and classification head will be updated.
+        """
+        self.encoder._freeze_parameters()
+
+    def get_input_embeddings(self) -> nn.Module:
+        return self.encoder.get_input_embeddings()
+
+    def set_input_embeddings(self, value: nn.Module):
+        self.encoder.set_input_embeddings(value)
+
+    @add_start_docstrings_to_model_forward(WHISPER_ENCODER_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=SequenceClassifierOutput, config_class=_CONFIG_FOR_DOC)
+    def forward(
+        self,
+        input_features: Optional[torch.LongTensor] = None,
+        head_mask: Optional[torch.Tensor] = None,
+        encoder_outputs: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,
+        labels: Optional[torch.LongTensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple[torch.Tensor], SequenceClassifierOutput]:
+        r"""
+        labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
+            Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
+            config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
+            `config.num_labels > 1` a classification loss is computed (Cross-Entropy).
+
+        Returns:
+
+        Example:
+
+        ```python
+        >>> import torch
+        >>> from transformers import AutoFeatureExtractor, WhisperForAudioClassification
+        >>> from datasets import load_dataset
+
+        >>> feature_extractor = AutoFeatureExtractor.from_pretrained("sanchit-gandhi/whisper-medium-fleurs-lang-id")
+        >>> model = WhisperForAudioClassification.from_pretrained("sanchit-gandhi/whisper-medium-fleurs-lang-id")
+
+        >>> ds = load_dataset("google/fleurs", "all", split="validation", streaming=True)
+        >>> sample = next(iter(ds))
+
+        >>> inputs = feature_extractor(
+        ...     sample["audio"]["array"], sampling_rate=sample["audio"]["sampling_rate"], return_tensors="pt"
+        ... )
+        >>> input_features = inputs.input_features
+
+        >>> with torch.no_grad():
+        ...     logits = model(input_features).logits
+
+        >>> predicted_class_ids = torch.argmax(logits).item()
+        >>> predicted_label = model.config.id2label[predicted_class_ids]
+        >>> predicted_label
+        'af_za'
+        ```"""
+
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        if encoder_outputs is None:
+            encoder_outputs = self.encoder(
+                input_features,
+                head_mask=head_mask,
+                output_attentions=output_attentions,
+                output_hidden_states=output_hidden_states,
+                return_dict=return_dict,
+            )
+
+        if self.config.use_weighted_layer_sum:
+            hidden_states = torch.stack(encoder_outputs, dim=1)
+            norm_weights = nn.functional.softmax(self.layer_weights, dim=-1)
+            hidden_states = (hidden_states * norm_weights.view(-1, 1, 1)).sum(dim=1)
+        else:
+            hidden_states = encoder_outputs[0]
+
+        hidden_states = self.projector(hidden_states)
+        pooled_output = hidden_states.mean(dim=1)
+
+        logits = self.classifier(pooled_output)
+
+        loss = None
+
+        if labels is not None:
+            loss_fct = CrossEntropyLoss()
+            # move labels to correct device to enable PP
+            labels = labels.to(logits.device)
+            loss = loss_fct(logits.view(-1, self.config.num_labels), labels.view(-1))
+
+        if not return_dict:
+            output = (logits,) + encoder_outputs[1:]
+            return ((loss,) + output) if loss is not None else output
+
+        return SequenceClassifierOutput(
+            loss=loss,
+            logits=logits,
+            hidden_states=encoder_outputs.hidden_states,
+            attentions=encoder_outputs.attentions,
+        )

modeling_whisper_mla.py

@@ -0,0 +1,64 @@
+"""WhisperMLA — Whisper with MLA (multi-head latent attention) converted decoder attention.
+
+Standalone trust_remote_code modeling: builds the MLA architecture structurally at init (the
+converted modules' weights then come from the checkpoint), so
+``AutoModelForSpeechSeq2Seq.from_pretrained(repo, trust_remote_code=True)`` just works.
+Pinned to transformers==4.46.x (the sibling vendored modeling targets it).
+
+Adapted from Whisper-MLA / MHA2MLA (Apache-2.0); see the repo NOTICE.
+"""
+import types
+from types import SimpleNamespace
+
+from .configuration_whisper_mla import WhisperMLAConfig
+from .modeling_whisper import WhisperAttention, WhisperForConditionalGeneration
+
+# NOTE: patch_func is imported directly (not only transitively) so the HF dynamic-module loader
+# copies it into the modules cache — it scans the ENTRY file's `from .x import y` lines.
+from .patch_func import svd_low_rank_approx as _ensure_patch_func_is_bundled  # noqa: F401
+from .patching_model_load import patch_model
+from .patching_whisper import (
+    custom_WhisperAttention_mla_forward,
+    custom_WhisperAttention_nosplit_forward,
+)
+
+_SCOPE_TARGETS = {
+    "decoder_self": frozenset({"decoder_self"}),
+    "plus_cross": frozenset({"decoder_self", "decoder_cross"}),
+    "all": frozenset({"decoder_self", "decoder_cross", "encoder_self"}),
+}
+
+
+def _structural_convert(model, config: WhisperMLAConfig) -> None:
+    """Rebuild the attention modules as MLA per the config. Runs on the freshly-initialized model
+    inside __init__ — the SVD of the (random) init weights only shapes the modules; real weights
+    arrive from the checkpoint via from_pretrained's state-dict load."""
+    if getattr(config, "num_key_value_heads", None) is None:
+        config.num_key_value_heads = config.encoder_attention_heads
+    preserve = config.mla_scheme == "partial_preservation"
+    args = SimpleNamespace(
+        rope_dim_for_mla=config.mla_rope_dim,
+        partial_rope_version={"uniform": "uniform", "2_norm": "2-norm"}[config.mla_selection],
+        low_rank=config.mla_low_rank,
+        svd_init_method="joint",
+        is_gqa2mha2mla=False,
+        uniform_start_point=0,
+        is_mla_from_scratch=False,
+        qk_tensor_path=None,
+        preserve_rope=preserve,
+        targets=set(_SCOPE_TARGETS[config.mla_scope]),
+    )
+    patch_model(model, config, args)
+    forward = custom_WhisperAttention_mla_forward if preserve else custom_WhisperAttention_nosplit_forward
+    for module in model.modules():
+        if isinstance(module, WhisperAttention) and hasattr(module, "kv_proj"):
+            module.forward = types.MethodType(forward, module)
+
+
+class WhisperMLAForConditionalGeneration(WhisperForConditionalGeneration):
+    config_class = WhisperMLAConfig
+
+    def __init__(self, config: WhisperMLAConfig):
+        super().__init__(config)
+        if config.mla_scheme != "none":
+            _structural_convert(self, config)

normalizer.json

@@ -0,0 +1,1742 @@
+{
+  "accessorise": "accessorize",
+  "accessorised": "accessorized",
+  "accessorises": "accessorizes",
+  "accessorising": "accessorizing",
+  "acclimatisation": "acclimatization",
+  "acclimatise": "acclimatize",
+  "acclimatised": "acclimatized",
+  "acclimatises": "acclimatizes",
+  "acclimatising": "acclimatizing",
+  "accoutrements": "accouterments",
+  "aeon": "eon",
+  "aeons": "eons",
+  "aerogramme": "aerogram",
+  "aerogrammes": "aerograms",
+  "aeroplane": "airplane",
+  "aeroplanes": "airplanes",
+  "aesthete": "esthete",
+  "aesthetes": "esthetes",
+  "aesthetic": "esthetic",
+  "aesthetically": "esthetically",
+  "aesthetics": "esthetics",
+  "aetiology": "etiology",
+  "ageing": "aging",
+  "aggrandisement": "aggrandizement",
+  "agonise": "agonize",
+  "agonised": "agonized",
+  "agonises": "agonizes",
+  "agonising": "agonizing",
+  "agonisingly": "agonizingly",
+  "almanack": "almanac",
+  "almanacks": "almanacs",
+  "aluminium": "aluminum",
+  "amortisable": "amortizable",
+  "amortisation": "amortization",
+  "amortisations": "amortizations",
+  "amortise": "amortize",
+  "amortised": "amortized",
+  "amortises": "amortizes",
+  "amortising": "amortizing",
+  "amphitheatre": "amphitheater",
+  "amphitheatres": "amphitheaters",
+  "anaemia": "anemia",
+  "anaemic": "anemic",
+  "anaesthesia": "anesthesia",
+  "anaesthetic": "anesthetic",
+  "anaesthetics": "anesthetics",
+  "anaesthetise": "anesthetize",
+  "anaesthetised": "anesthetized",
+  "anaesthetises": "anesthetizes",
+  "anaesthetising": "anesthetizing",
+  "anaesthetist": "anesthetist",
+  "anaesthetists": "anesthetists",
+  "anaesthetize": "anesthetize",
+  "anaesthetized": "anesthetized",
+  "anaesthetizes": "anesthetizes",
+  "anaesthetizing": "anesthetizing",
+  "analogue": "analog",
+  "analogues": "analogs",
+  "analyse": "analyze",
+  "analysed": "analyzed",
+  "analyses": "analyzes",
+  "analysing": "analyzing",
+  "anglicise": "anglicize",
+  "anglicised": "anglicized",
+  "anglicises": "anglicizes",
+  "anglicising": "anglicizing",
+  "annualised": "annualized",
+  "antagonise": "antagonize",
+  "antagonised": "antagonized",
+  "antagonises": "antagonizes",
+  "antagonising": "antagonizing",
+  "apologise": "apologize",
+  "apologised": "apologized",
+  "apologises": "apologizes",
+  "apologising": "apologizing",
+  "appal": "appall",
+  "appals": "appalls",
+  "appetiser": "appetizer",
+  "appetisers": "appetizers",
+  "appetising": "appetizing",
+  "appetisingly": "appetizingly",
+  "arbour": "arbor",
+  "arbours": "arbors",
+  "archaeologically": "archeologically",
+  "archaeologist": "archeologist",
+  "archaeologists": "archeologists",
+  "archaeology": "archeology</span>",
+  "archeological": "archaeological",
+  "ardour": "ardor",
+  "armour": "armor",
+  "armoured": "armored",
+  "armourer": "armorer",
+  "armourers": "armorers",
+  "armouries": "armories",
+  "armoury": "armory",
+  "artefact": "artifact",
+  "artefacts": "artifacts",
+  "authorise": "authorize",
+  "authorised": "authorized",
+  "authorises": "authorizes",
+  "authorising": "authorizing",
+  "axe": "ax",
+  "backpedalled": "backpedaled",
+  "backpedalling": "backpedaling",
+  "bannister": "banister",
+  "bannisters": "banisters",
+  "baptise": "baptize",
+  "baptised": "baptized",
+  "baptises": "baptizes",
+  "baptising": "baptizing",
+  "bastardise": "bastardize",
+  "bastardised": "bastardized",
+  "bastardises": "bastardizes",
+  "bastardising": "bastardizing",
+  "battleax": "battleaxe",
+  "baulk": "balk",
+  "baulked": "balked",
+  "baulking": "balking",
+  "baulks": "balks",
+  "bedevilled": "bedeviled",
+  "bedevilling": "bedeviling",
+  "behaviour": "behavior",
+  "behavioural": "behavioral",
+  "behaviourism": "behaviorism",
+  "behaviourist": "behaviorist",
+  "behaviourists": "behaviorists",
+  "behaviours": "behaviors",
+  "behove": "behoove",
+  "behoved": "behooved",
+  "behoves": "behooves",
+  "bejewelled": "bejeweled",
+  "belabour": "belabor",
+  "belaboured": "belabored",
+  "belabouring": "belaboring",
+  "belabours": "belabors",
+  "bevelled": "beveled",
+  "bevvies": "bevies",
+  "bevvy": "bevy",
+  "biassed": "biased",
+  "biassing": "biasing",
+  "bingeing": "binging",
+  "bougainvillaea": "bougainvillea",
+  "bougainvillaeas": "bougainvilleas",
+  "bowdlerise": "bowdlerize",
+  "bowdlerised": "bowdlerized",
+  "bowdlerises": "bowdlerizes",
+  "bowdlerising": "bowdlerizing",
+  "breathalyse": "breathalyze",
+  "breathalysed": "breathalyzed",
+  "breathalyser": "breathalyzer",
+  "breathalysers": "breathalyzers",
+  "breathalyses": "breathalyzes",
+  "breathalysing": "breathalyzing",
+  "brutalise": "brutalize",
+  "brutalised": "brutalized",
+  "brutalises": "brutalizes",
+  "brutalising": "brutalizing",
+  "busses": "buses",
+  "bussing": "busing",
+  "caesarean": "cesarean",
+  "caesareans": "cesareans",
+  "calibre": "caliber",
+  "calibres": "calibers",
+  "calliper": "caliper",
+  "callipers": "calipers",
+  "callisthenics": "calisthenics",
+  "canalise": "canalize",
+  "canalised": "canalized",
+  "canalises": "canalizes",
+  "canalising": "canalizing",
+  "cancelation": "cancellation",
+  "cancelations": "cancellations",
+  "cancelled": "canceled",
+  "cancelling": "canceling",
+  "candour": "candor",
+  "cannibalise": "cannibalize",
+  "cannibalised": "cannibalized",
+  "cannibalises": "cannibalizes",
+  "cannibalising": "cannibalizing",
+  "canonise": "canonize",
+  "canonised": "canonized",
+  "canonises": "canonizes",
+  "canonising": "canonizing",
+  "capitalise": "capitalize",
+  "capitalised": "capitalized",
+  "capitalises": "capitalizes",
+  "capitalising": "capitalizing",
+  "caramelise": "caramelize",
+  "caramelised": "caramelized",
+  "caramelises": "caramelizes",
+  "caramelising": "caramelizing",
+  "carbonise": "carbonize",
+  "carbonised": "carbonized",
+  "carbonises": "carbonizes",
+  "carbonising": "carbonizing",
+  "carolled": "caroled",
+  "carolling": "caroling",
+  "catalogue": "catalog",
+  "catalogued": "cataloged",
+  "catalogues": "catalogs",
+  "cataloguing": "cataloging",
+  "catalyse": "catalyze",
+  "catalysed": "catalyzed",
+  "catalyses": "catalyzes",
+  "catalysing": "catalyzing",
+  "categorise": "categorize",
+  "categorised": "categorized",
+  "categorises": "categorizes",
+  "categorising": "categorizing",
+  "cauterise": "cauterize",
+  "cauterised": "cauterized",
+  "cauterises": "cauterizes",
+  "cauterising": "cauterizing",
+  "cavilled": "caviled",
+  "cavilling": "caviling",
+  "centigramme": "centigram",
+  "centigrammes": "centigrams",
+  "centilitre": "centiliter",
+  "centilitres": "centiliters",
+  "centimetre": "centimeter",
+  "centimetres": "centimeters",
+  "centralise": "centralize",
+  "centralised": "centralized",
+  "centralises": "centralizes",
+  "centralising": "centralizing",
+  "centre": "center",
+  "centred": "centered",
+  "centrefold": "centerfold",
+  "centrefolds": "centerfolds",
+  "centrepiece": "centerpiece",
+  "centrepieces": "centerpieces",
+  "centres": "centers",
+  "channelled": "channeled",
+  "channelling": "channeling",
+  "characterise": "characterize",
+  "characterised": "characterized",
+  "characterises": "characterizes",
+  "characterising": "characterizing",
+  "cheque": "check",
+  "chequebook": "checkbook",
+  "chequebooks": "checkbooks",
+  "chequered": "checkered",
+  "cheques": "checks",
+  "chilli": "chili",
+  "chimaera": "chimera",
+  "chimaeras": "chimeras",
+  "chiselled": "chiseled",
+  "chiselling": "chiseling",
+  "circularise": "circularize",
+  "circularised": "circularized",
+  "circularises": "circularizes",
+  "circularising": "circularizing",
+  "civilise": "civilize",
+  "civilised": "civilized",
+  "civilises": "civilizes",
+  "civilising": "civilizing",
+  "clamour": "clamor",
+  "clamoured": "clamored",
+  "clamouring": "clamoring",
+  "clamours": "clamors",
+  "clangour": "clangor",
+  "clarinettist": "clarinetist",
+  "clarinettists": "clarinetists",
+  "collectivise": "collectivize",
+  "collectivised": "collectivized",
+  "collectivises": "collectivizes",
+  "collectivising": "collectivizing",
+  "colonisation": "colonization",
+  "colonise": "colonize",
+  "colonised": "colonized",
+  "coloniser": "colonizer",
+  "colonisers": "colonizers",
+  "colonises": "colonizes",
+  "colonising": "colonizing",
+  "colour": "color",
+  "colourant": "colorant",
+  "colourants": "colorants",
+  "coloured": "colored",
+  "coloureds": "coloreds",
+  "colourful": "colorful",
+  "colourfully": "colorfully",
+  "colouring": "coloring",
+  "colourize": "colorize",
+  "colourized": "colorized",
+  "colourizes": "colorizes",
+  "colourizing": "colorizing",
+  "colourless": "colorless",
+  "colours": "colors",
+  "commercialise": "commercialize",
+  "commercialised": "commercialized",
+  "commercialises": "commercializes",
+  "commercialising": "commercializing",
+  "compartmentalise": "compartmentalize",
+  "compartmentalised": "compartmentalized",
+  "compartmentalises": "compartmentalizes",
+  "compartmentalising": "compartmentalizing",
+  "computerise": "computerize",
+  "computerised": "computerized",
+  "computerises": "computerizes",
+  "computerising": "computerizing",
+  "conceptualise": "conceptualize",
+  "conceptualised": "conceptualized",
+  "conceptualises": "conceptualizes",
+  "conceptualising": "conceptualizing",
+  "connexion": "connection",
+  "connexions": "connections",
+  "contextualise": "contextualize",
+  "contextualised": "contextualized",
+  "contextualises": "contextualizes",
+  "contextualising": "contextualizing",
+  "cosier": "cozier",
+  "cosies": "cozies",
+  "cosiest": "coziest",
+  "cosily": "cozily",
+  "cosiness": "coziness",
+  "cosy": "cozy",
+  "councillor": "councilor",
+  "councillors": "councilors",
+  "counselled": "counseled",
+  "counselling": "counseling",
+  "counsellor": "counselor",
+  "counsellors": "counselors",
+  "crenelated": "crenellated",
+  "criminalise": "criminalize",
+  "criminalised": "criminalized",
+  "criminalises": "criminalizes",
+  "criminalising": "criminalizing",
+  "criticise": "criticize",
+  "criticised": "criticized",
+  "criticises": "criticizes",
+  "criticising": "criticizing",
+  "crueller": "crueler",
+  "cruellest": "cruelest",
+  "crystallisation": "crystallization",
+  "crystallise": "crystallize",
+  "crystallised": "crystallized",
+  "crystallises": "crystallizes",
+  "crystallising": "crystallizing",
+  "cudgelled": "cudgeled",
+  "cudgelling": "cudgeling",
+  "customise": "customize",
+  "customised": "customized",
+  "customises": "customizes",
+  "customising": "customizing",
+  "cypher": "cipher",
+  "cyphers": "ciphers",
+  "decentralisation": "decentralization",
+  "decentralise": "decentralize",
+  "decentralised": "decentralized",
+  "decentralises": "decentralizes",
+  "decentralising": "decentralizing",
+  "decriminalisation": "decriminalization",
+  "decriminalise": "decriminalize",
+  "decriminalised": "decriminalized",
+  "decriminalises": "decriminalizes",
+  "decriminalising": "decriminalizing",
+  "defence": "defense",
+  "defenceless": "defenseless",
+  "defences": "defenses",
+  "dehumanisation": "dehumanization",
+  "dehumanise": "dehumanize",
+  "dehumanised": "dehumanized",
+  "dehumanises": "dehumanizes",
+  "dehumanising": "dehumanizing",
+  "demeanour": "demeanor",
+  "demilitarisation": "demilitarization",
+  "demilitarise": "demilitarize",
+  "demilitarised": "demilitarized",
+  "demilitarises": "demilitarizes",
+  "demilitarising": "demilitarizing",
+  "demobilisation": "demobilization",
+  "demobilise": "demobilize",
+  "demobilised": "demobilized",
+  "demobilises": "demobilizes",
+  "demobilising": "demobilizing",
+  "democratisation": "democratization",
+  "democratise": "democratize",
+  "democratised": "democratized",
+  "democratises": "democratizes",
+  "democratising": "democratizing",
+  "demonise": "demonize",
+  "demonised": "demonized",
+  "demonises": "demonizes",
+  "demonising": "demonizing",
+  "demoralisation": "demoralization",
+  "demoralise": "demoralize",
+  "demoralised": "demoralized",
+  "demoralises": "demoralizes",
+  "demoralising": "demoralizing",
+  "denationalisation": "denationalization",
+  "denationalise": "denationalize",
+  "denationalised": "denationalized",
+  "denationalises": "denationalizes",
+  "denationalising": "denationalizing",
+  "deodorise": "deodorize",
+  "deodorised": "deodorized",
+  "deodorises": "deodorizes",
+  "deodorising": "deodorizing",
+  "depersonalise": "depersonalize",
+  "depersonalised": "depersonalized",
+  "depersonalises": "depersonalizes",
+  "depersonalising": "depersonalizing",
+  "deputise": "deputize",
+  "deputised": "deputized",
+  "deputises": "deputizes",
+  "deputising": "deputizing",
+  "desensitisation": "desensitization",
+  "desensitise": "desensitize",
+  "desensitised": "desensitized",
+  "desensitises": "desensitizes",
+  "desensitising": "desensitizing",
+  "destabilisation": "destabilization",
+  "destabilise": "destabilize",
+  "destabilised": "destabilized",
+  "destabilises": "destabilizes",
+  "destabilising": "destabilizing",
+  "dialled": "dialed",
+  "dialling": "dialing",
+  "dialogue": "dialog",
+  "dialogues": "dialogs",
+  "diarrhoea": "diarrhea",
+  "digitise": "digitize",
+  "digitised": "digitized",
+  "digitises": "digitizes",
+  "digitising": "digitizing",
+  "disc": "disk",
+  "discolour": "discolor",
+  "discoloured": "discolored",
+  "discolouring": "discoloring",
+  "discolours": "discolors",
+  "discs": "disks",
+  "disembowelled": "disemboweled",
+  "disembowelling": "disemboweling",
+  "disfavour": "disfavor",
+  "dishevelled": "disheveled",
+  "dishonour": "dishonor",
+  "dishonourable": "dishonorable",
+  "dishonourably": "dishonorably",
+  "dishonoured": "dishonored",
+  "dishonouring": "dishonoring",
+  "dishonours": "dishonors",
+  "disorganisation": "disorganization",
+  "disorganised": "disorganized",
+  "distil": "distill",
+  "distils": "distills",
+  "dramatisation": "dramatization",
+  "dramatisations": "dramatizations",
+  "dramatise": "dramatize",
+  "dramatised": "dramatized",
+  "dramatises": "dramatizes",
+  "dramatising": "dramatizing",
+  "draught": "draft",
+  "draughtboard": "draftboard",
+  "draughtboards": "draftboards",
+  "draughtier": "draftier",
+  "draughtiest": "draftiest",
+  "draughts": "drafts",
+  "draughtsman": "draftsman",
+  "draughtsmanship": "draftsmanship",
+  "draughtsmen": "draftsmen",
+  "draughtswoman": "draftswoman",
+  "draughtswomen": "draftswomen",
+  "draughty": "drafty",
+  "drivelled": "driveled",
+  "drivelling": "driveling",
+  "duelled": "dueled",
+  "duelling": "dueling",
+  "economise": "economize",
+  "economised": "economized",
+  "economises": "economizes",
+  "economising": "economizing",
+  "editorialise": "editorialize",
+  "editorialised": "editorialized",
+  "editorialises": "editorializes",
+  "editorialising": "editorializing",
+  "edoema": "edema",
+  "empathise": "empathize",
+  "empathised": "empathized",
+  "empathises": "empathizes",
+  "empathising": "empathizing",
+  "emphasise": "emphasize",
+  "emphasised": "emphasized",
+  "emphasises": "emphasizes",
+  "emphasising": "emphasizing",
+  "enamelled": "enameled",
+  "enamelling": "enameling",
+  "enamoured": "enamored",
+  "encyclopaedia": "encyclopedia",
+  "encyclopaedias": "encyclopedias",
+  "encyclopaedic": "encyclopedic",
+  "endeavour": "endeavor",
+  "endeavoured": "endeavored",
+  "endeavouring": "endeavoring",
+  "endeavours": "endeavors",
+  "energise": "energize",
+  "energised": "energized",
+  "energises": "energizes",
+  "energising": "energizing",
+  "enrol": "enroll",
+  "enrols": "enrolls",
+  "enthral": "enthrall",
+  "enthrals": "enthralls",
+  "epaulette": "epaulet",
+  "epaulettes": "epaulets",
+  "epicentre": "epicenter",
+  "epicentres": "epicenters",
+  "epilogue": "epilog",
+  "epilogues": "epilogs",
+  "epitomise": "epitomize",
+  "epitomised": "epitomized",
+  "epitomises": "epitomizes",
+  "epitomising": "epitomizing",
+  "equalisation": "equalization",
+  "equalise": "equalize",
+  "equalised": "equalized",
+  "equaliser": "equalizer",
+  "equalisers": "equalizers",
+  "equalises": "equalizes",
+  "equalising": "equalizing",
+  "eulogise": "eulogize",
+  "eulogised": "eulogized",
+  "eulogises": "eulogizes",
+  "eulogising": "eulogizing",
+  "evangelise": "evangelize",
+  "evangelised": "evangelized",
+  "evangelises": "evangelizes",
+  "evangelising": "evangelizing",
+  "exorcise": "exorcize",
+  "exorcised": "exorcized",
+  "exorcises": "exorcizes",
+  "exorcising": "exorcizing",
+  "extemporisation": "extemporization",
+  "extemporise": "extemporize",
+  "extemporised": "extemporized",
+  "extemporises": "extemporizes",
+  "extemporising": "extemporizing",
+  "externalisation": "externalization",
+  "externalisations": "externalizations",
+  "externalise": "externalize",
+  "externalised": "externalized",
+  "externalises": "externalizes",
+  "externalising": "externalizing",
+  "factorise": "factorize",
+  "factorised": "factorized",
+  "factorises": "factorizes",
+  "factorising": "factorizing",
+  "faecal": "fecal",
+  "faeces": "feces",
+  "familiarisation": "familiarization",
+  "familiarise": "familiarize",
+  "familiarised": "familiarized",
+  "familiarises": "familiarizes",
+  "familiarising": "familiarizing",
+  "fantasise": "fantasize",
+  "fantasised": "fantasized",
+  "fantasises": "fantasizes",
+  "fantasising": "fantasizing",
+  "favour": "favor",
+  "favourable": "favorable",
+  "favourably": "favorably",
+  "favoured": "favored",
+  "favouring": "favoring",
+  "favourite": "favorite",
+  "favourites": "favorites",
+  "favouritism": "favoritism",
+  "favours": "favors",
+  "feminise": "feminize",
+  "feminised": "feminized",
+  "feminises": "feminizes",
+  "feminising": "feminizing",
+  "fertilisation": "fertilization",
+  "fertilise": "fertilize",
+  "fertilised": "fertilized",
+  "fertiliser": "fertilizer",
+  "fertilisers": "fertilizers",
+  "fertilises": "fertilizes",
+  "fertilising": "fertilizing",
+  "fervour": "fervor",
+  "fibre": "fiber",
+  "fibreglass": "fiberglass",
+  "fibres": "fibers",
+  "fictionalisation": "fictionalization",
+  "fictionalisations": "fictionalizations",
+  "fictionalise": "fictionalize",
+  "fictionalised": "fictionalized",
+  "fictionalises": "fictionalizes",
+  "fictionalising": "fictionalizing",
+  "fillet": "filet",
+  "filleted": "fileted",
+  "filleting": "fileting",
+  "fillets": "filets",
+  "finalisation": "finalization",
+  "finalise": "finalize",
+  "finalised": "finalized",
+  "finalises": "finalizes",
+  "finalising": "finalizing",
+  "flautist": "flutist",
+  "flautists": "flutists",
+  "flavour": "flavor",
+  "flavoured": "flavored",
+  "flavouring": "flavoring",
+  "flavourings": "flavorings",
+  "flavourless": "flavorless",
+  "flavours": "flavors",
+  "flavoursome": "flavorsome",
+  "flyer / flier": "flier / flyer",
+  "foetal": "fetal",
+  "foetid": "fetid",
+  "foetus": "fetus",
+  "foetuses": "fetuses",
+  "formalisation": "formalization",
+  "formalise": "formalize",
+  "formalised": "formalized",
+  "formalises": "formalizes",
+  "formalising": "formalizing",
+  "fossilisation": "fossilization",
+  "fossilise": "fossilize",
+  "fossilised": "fossilized",
+  "fossilises": "fossilizes",
+  "fossilising": "fossilizing",
+  "fraternisation": "fraternization",
+  "fraternise": "fraternize",
+  "fraternised": "fraternized",
+  "fraternises": "fraternizes",
+  "fraternising": "fraternizing",
+  "fulfil": "fulfill",
+  "fulfilment": "fulfillment",
+  "fulfils": "fulfills",
+  "funnelled": "funneled",
+  "funnelling": "funneling",
+  "gage": "gauge",
+  "gaged": "gauged",
+  "gages": "gauges",
+  "gaging": "gauging",
+  "galvanise": "galvanize",
+  "galvanised": "galvanized",
+  "galvanises": "galvanizes",
+  "galvanising": "galvanizing",
+  "gambolled": "gamboled",
+  "gambolling": "gamboling",
+  "gaol": "jail",
+  "gaolbird": "jailbird",
+  "gaolbirds": "jailbirds",
+  "gaolbreak": "jailbreak",
+  "gaolbreaks": "jailbreaks",
+  "gaoled": "jailed",
+  "gaoler": "jailer",
+  "gaolers": "jailers",
+  "gaoling": "jailing",
+  "gaols": "jails",
+  "gasses": "gases",
+  "generalisation": "generalization",
+  "generalisations": "generalizations",
+  "generalise": "generalize",
+  "generalised": "generalized",
+  "generalises": "generalizes",
+  "generalising": "generalizing",
+  "ghettoise": "ghettoize",
+  "ghettoised": "ghettoized",
+  "ghettoises": "ghettoizes",
+  "ghettoising": "ghettoizing",
+  "gipsies": "gypsies",
+  "glamor": "glamour",
+  "glamorise": "glamorize",
+  "glamorised": "glamorized",
+  "glamorises": "glamorizes",
+  "glamorising": "glamorizing",
+  "globalisation": "globalization",
+  "globalise": "globalize",
+  "globalised": "globalized",
+  "globalises": "globalizes",
+  "globalising": "globalizing",
+  "glueing": "gluing",
+  "goitre": "goiter",
+  "goitres": "goiters",
+  "gonorrhoea": "gonorrhea",
+  "gramme": "gram",
+  "grammes": "grams",
+  "gravelled": "graveled",
+  "grey": "gray",
+  "greyed": "grayed",
+  "greying": "graying",
+  "greyish": "grayish",
+  "greyness": "grayness",
+  "greys": "grays",
+  "grovelled": "groveled",
+  "grovelling": "groveling",
+  "groyne": "groin",
+  "groynes": "groins",
+  "gruelling": "grueling",
+  "gruellingly": "gruelingly",
+  "gryphon": "griffin",
+  "gryphons": "griffins",
+  "gynaecological": "gynecological",
+  "gynaecologist": "gynecologist",
+  "gynaecologists": "gynecologists",
+  "gynaecology": "gynecology",
+  "haematological": "hematological",
+  "haematologist": "hematologist",
+  "haematologists": "hematologists",
+  "haematology": "hematology",
+  "haemoglobin": "hemoglobin",
+  "haemophilia": "hemophilia",
+  "haemophiliac": "hemophiliac",
+  "haemophiliacs": "hemophiliacs",
+  "haemorrhage": "hemorrhage",
+  "haemorrhaged": "hemorrhaged",
+  "haemorrhages": "hemorrhages",
+  "haemorrhaging": "hemorrhaging",
+  "haemorrhoids": "hemorrhoids",
+  "harbour": "harbor",
+  "harboured": "harbored",
+  "harbouring": "harboring",
+  "harbours": "harbors",
+  "harmonisation": "harmonization",
+  "harmonise": "harmonize",
+  "harmonised": "harmonized",
+  "harmonises": "harmonizes",
+  "harmonising": "harmonizing",
+  "homoeopath": "homeopath",
+  "homoeopathic": "homeopathic",
+  "homoeopaths": "homeopaths",
+  "homoeopathy": "homeopathy",
+  "homogenise": "homogenize",
+  "homogenised": "homogenized",
+  "homogenises": "homogenizes",
+  "homogenising": "homogenizing",
+  "honour": "honor",
+  "honourable": "honorable",
+  "honourably": "honorably",
+  "honoured": "honored",
+  "honouring": "honoring",
+  "honours": "honors",
+  "hospitalisation": "hospitalization",
+  "hospitalise": "hospitalize",
+  "hospitalised": "hospitalized",
+  "hospitalises": "hospitalizes",
+  "hospitalising": "hospitalizing",
+  "humanise": "humanize",
+  "humanised": "humanized",
+  "humanises": "humanizes",
+  "humanising": "humanizing",
+  "humour": "humor",
+  "humoured": "humored",
+  "humouring": "humoring",
+  "humourless": "humorless",
+  "humours": "humors",
+  "hybridise": "hybridize",
+  "hybridised": "hybridized",
+  "hybridises": "hybridizes",
+  "hybridising": "hybridizing",
+  "hypnotise": "hypnotize",
+  "hypnotised": "hypnotized",
+  "hypnotises": "hypnotizes",
+  "hypnotising": "hypnotizing",
+  "hypothesise": "hypothesize",
+  "hypothesised": "hypothesized",
+  "hypothesises": "hypothesizes",
+  "hypothesising": "hypothesizing",
+  "idealisation": "idealization",
+  "idealise": "idealize",
+  "idealised": "idealized",
+  "idealises": "idealizes",
+  "idealising": "idealizing",
+  "idolise": "idolize",
+  "idolised": "idolized",
+  "idolises": "idolizes",
+  "idolising": "idolizing",
+  "immobilisation": "immobilization",
+  "immobilise": "immobilize",
+  "immobilised": "immobilized",
+  "immobiliser": "immobilizer",
+  "immobilisers": "immobilizers",
+  "immobilises": "immobilizes",
+  "immobilising": "immobilizing",
+  "immortalise": "immortalize",
+  "immortalised": "immortalized",
+  "immortalises": "immortalizes",
+  "immortalising": "immortalizing",
+  "immunisation": "immunization",
+  "immunise": "immunize",
+  "immunised": "immunized",
+  "immunises": "immunizes",
+  "immunising": "immunizing",
+  "impanelled": "impaneled",
+  "impanelling": "impaneling",
+  "imperilled": "imperiled",
+  "imperilling": "imperiling",
+  "individualise": "individualize",
+  "individualised": "individualized",
+  "individualises": "individualizes",
+  "individualising": "individualizing",
+  "industrialise": "industrialize",
+  "industrialised": "industrialized",
+  "industrialises": "industrializes",
+  "industrialising": "industrializing",
+  "inflexion": "inflection",
+  "inflexions": "inflections",
+  "initialise": "initialize",
+  "initialised": "initialized",
+  "initialises": "initializes",
+  "initialising": "initializing",
+  "initialled": "initialed",
+  "initialling": "initialing",
+  "instal": "install",
+  "instalment": "installment",
+  "instalments": "installments",
+  "instals": "installs",
+  "instil": "instill",
+  "instils": "instills",
+  "institutionalisation": "institutionalization",
+  "institutionalise": "institutionalize",
+  "institutionalised": "institutionalized",
+  "institutionalises": "institutionalizes",
+  "institutionalising": "institutionalizing",
+  "intellectualise": "intellectualize",
+  "intellectualised": "intellectualized",
+  "intellectualises": "intellectualizes",
+  "intellectualising": "intellectualizing",
+  "internalisation": "internalization",
+  "internalise": "internalize",
+  "internalised": "internalized",
+  "internalises": "internalizes",
+  "internalising": "internalizing",
+  "internationalisation": "internationalization",
+  "internationalise": "internationalize",
+  "internationalised": "internationalized",
+  "internationalises": "internationalizes",
+  "internationalising": "internationalizing",
+  "ionisation": "ionization",
+  "ionise": "ionize",
+  "ionised": "ionized",
+  "ioniser": "ionizer",
+  "ionisers": "ionizers",
+  "ionises": "ionizes",
+  "ionising": "ionizing",
+  "italicise": "italicize",
+  "italicised": "italicized",
+  "italicises": "italicizes",
+  "italicising": "italicizing",
+  "itemise": "itemize",
+  "itemised": "itemized",
+  "itemises": "itemizes",
+  "itemising": "itemizing",
+  "jeopardise": "jeopardize",
+  "jeopardised": "jeopardized",
+  "jeopardises": "jeopardizes",
+  "jeopardising": "jeopardizing",
+  "jewelled": "jeweled",
+  "jeweller": "jeweler",
+  "jewellers": "jewelers",
+  "jewellery": "jewelry",
+  "judgement": "judgment",
+  "kilogramme": "kilogram",
+  "kilogrammes": "kilograms",
+  "kilometre": "kilometer",
+  "kilometres": "kilometers",
+  "labelled": "labeled",
+  "labelling": "labeling",
+  "labour": "labor",
+  "laboured": "labored",
+  "labourer": "laborer",
+  "labourers": "laborers",
+  "labouring": "laboring",
+  "labours": "labors",
+  "lacklustre": "lackluster",
+  "legalisation": "legalization",
+  "legalise": "legalize",
+  "legalised": "legalized",
+  "legalises": "legalizes",
+  "legalising": "legalizing",
+  "legitimise": "legitimize",
+  "legitimised": "legitimized",
+  "legitimises": "legitimizes",
+  "legitimising": "legitimizing",
+  "leukaemia": "leukemia",
+  "levelled": "leveled",
+  "leveller": "leveler",
+  "levellers": "levelers",
+  "levelling": "leveling",
+  "libelled": "libeled",
+  "libelling": "libeling",
+  "libellous": "libelous",
+  "liberalisation": "liberalization",
+  "liberalise": "liberalize",
+  "liberalised": "liberalized",
+  "liberalises": "liberalizes",
+  "liberalising": "liberalizing",
+  "licence": "license",
+  "licenced": "licensed",
+  "licences": "licenses",
+  "licencing": "licensing",
+  "likeable": "likable",
+  "lionisation": "lionization",
+  "lionise": "lionize",
+  "lionised": "lionized",
+  "lionises": "lionizes",
+  "lionising": "lionizing",
+  "liquidise": "liquidize",
+  "liquidised": "liquidized",
+  "liquidiser": "liquidizer",
+  "liquidisers": "liquidizers",
+  "liquidises": "liquidizes",
+  "liquidising": "liquidizing",
+  "litre": "liter",
+  "litres": "liters",
+  "localise": "localize",
+  "localised": "localized",
+  "localises": "localizes",
+  "localising": "localizing",
+  "louvre": "louver",
+  "louvred": "louvered",
+  "louvres": "louvers",
+  "lustre": "luster",
+  "magnetise": "magnetize",
+  "magnetised": "magnetized",
+  "magnetises": "magnetizes",
+  "magnetising": "magnetizing",
+  "manoeuvrability": "maneuverability",
+  "manoeuvrable": "maneuverable",
+  "manoeuvre": "maneuver",
+  "manoeuvred": "maneuvered",
+  "manoeuvres": "maneuvers",
+  "manoeuvring": "maneuvering",
+  "manoeuvrings": "maneuverings",
+  "marginalisation": "marginalization",
+  "marginalise": "marginalize",
+  "marginalised": "marginalized",
+  "marginalises": "marginalizes",
+  "marginalising": "marginalizing",
+  "marshalled": "marshaled",
+  "marshalling": "marshaling",
+  "marvelled": "marveled",
+  "marvelling": "marveling",
+  "marvellous": "marvelous",
+  "marvellously": "marvelously",
+  "materialisation": "materialization",
+  "materialise": "materialize",
+  "materialised": "materialized",
+  "materialises": "materializes",
+  "materialising": "materializing",
+  "maximisation": "maximization",
+  "maximise": "maximize",
+  "maximised": "maximized",
+  "maximises": "maximizes",
+  "maximising": "maximizing",
+  "meagre": "meager",
+  "mechanisation": "mechanization",
+  "mechanise": "mechanize",
+  "mechanised": "mechanized",
+  "mechanises": "mechanizes",
+  "mechanising": "mechanizing",
+  "mediaeval": "medieval",
+  "memorialise": "memorialize",
+  "memorialised": "memorialized",
+  "memorialises": "memorializes",
+  "memorialising": "memorializing",
+  "memorise": "memorize",
+  "memorised": "memorized",
+  "memorises": "memorizes",
+  "memorising": "memorizing",
+  "mesmerise": "mesmerize",
+  "mesmerised": "mesmerized",
+  "mesmerises": "mesmerizes",
+  "mesmerising": "mesmerizing",
+  "metabolise": "metabolize",
+  "metabolised": "metabolized",
+  "metabolises": "metabolizes",
+  "metabolising": "metabolizing",
+  "metre": "meter",
+  "metres": "meters",
+  "mhm": "hmm",
+  "micrometre": "micrometer",
+  "micrometres": "micrometers",
+  "militarise": "militarize",
+  "militarised": "militarized",
+  "militarises": "militarizes",
+  "militarising": "militarizing",
+  "milligramme": "milligram",
+  "milligrammes": "milligrams",
+  "millilitre": "milliliter",
+  "millilitres": "milliliters",
+  "millimetre": "millimeter",
+  "millimetres": "millimeters",
+  "miniaturisation": "miniaturization",
+  "miniaturise": "miniaturize",
+  "miniaturised": "miniaturized",
+  "miniaturises": "miniaturizes",
+  "miniaturising": "miniaturizing",
+  "minibusses": "minibuses",
+  "minimise": "minimize",
+  "minimised": "minimized",
+  "minimises": "minimizes",
+  "minimising": "minimizing",
+  "misbehaviour": "misbehavior",
+  "misdemeanour": "misdemeanor",
+  "misdemeanours": "misdemeanors",
+  "misspelt": "misspelled",
+  "mitre": "miter",
+  "mitres": "miters",
+  "mm": "hmm",
+  "mmm": "hmm",
+  "mobilisation": "mobilization",
+  "mobilise": "mobilize",
+  "mobilised": "mobilized",
+  "mobilises": "mobilizes",
+  "mobilising": "mobilizing",
+  "modelled": "modeled",
+  "modeller": "modeler",
+  "modellers": "modelers",
+  "modelling": "modeling",
+  "modernise": "modernize",
+  "modernised": "modernized",
+  "modernises": "modernizes",
+  "modernising": "modernizing",
+  "moisturise": "moisturize",
+  "moisturised": "moisturized",
+  "moisturiser": "moisturizer",
+  "moisturisers": "moisturizers",
+  "moisturises": "moisturizes",
+  "moisturising": "moisturizing",
+  "monologue": "monolog",
+  "monologues": "monologs",
+  "monopolisation": "monopolization",
+  "monopolise": "monopolize",
+  "monopolised": "monopolized",
+  "monopolises": "monopolizes",
+  "monopolising": "monopolizing",
+  "moralise": "moralize",
+  "moralised": "moralized",
+  "moralises": "moralizes",
+  "moralising": "moralizing",
+  "motorised": "motorized",
+  "mould": "mold",
+  "moulded": "molded",
+  "moulder": "molder",
+  "mouldered": "moldered",
+  "mouldering": "moldering",
+  "moulders": "molders",
+  "mouldier": "moldier",
+  "mouldiest": "moldiest",
+  "moulding": "molding",
+  "mouldings": "moldings",
+  "moulds": "molds",
+  "mouldy": "moldy",
+  "moult": "molt",
+  "moulted": "molted",
+  "moulting": "molting",
+  "moults": "molts",
+  "moustache": "mustache",
+  "moustached": "mustached",
+  "moustaches": "mustaches",
+  "moustachioed": "mustachioed",
+  "multicoloured": "multicolored",
+  "nationalisation": "nationalization",
+  "nationalisations": "nationalizations",
+  "nationalise": "nationalize",
+  "nationalised": "nationalized",
+  "nationalises": "nationalizes",
+  "nationalising": "nationalizing",
+  "naturalisation": "naturalization",
+  "naturalise": "naturalize",
+  "naturalised": "naturalized",
+  "naturalises": "naturalizes",
+  "naturalising": "naturalizing",
+  "neighbour": "neighbor",
+  "neighbourhood": "neighborhood",
+  "neighbourhoods": "neighborhoods",
+  "neighbouring": "neighboring",
+  "neighbourliness": "neighborliness",
+  "neighbourly": "neighborly",
+  "neighbours": "neighbors",
+  "neutralisation": "neutralization",
+  "neutralise": "neutralize",
+  "neutralised": "neutralized",
+  "neutralises": "neutralizes",
+  "neutralising": "neutralizing",
+  "normalisation": "normalization",
+  "normalise": "normalize",
+  "normalised": "normalized",
+  "normalises": "normalizes",
+  "normalising": "normalizing",
+  "odour": "odor",
+  "odourless": "odorless",
+  "odours": "odors",
+  "oesophagus": "esophagus",
+  "oesophaguses": "esophaguses",
+  "oestrogen": "estrogen",
+  "offence": "offense",
+  "offences": "offenses",
+  "omelette": "omelet",
+  "omelettes": "omelets",
+  "optimise": "optimize",
+  "optimised": "optimized",
+  "optimises": "optimizes",
+  "optimising": "optimizing",
+  "organisation": "organization",
+  "organisational": "organizational",
+  "organisations": "organizations",
+  "organise": "organize",
+  "organised": "organized",
+  "organiser": "organizer",
+  "organisers": "organizers",
+  "organises": "organizes",
+  "organising": "organizing",
+  "orthopaedic": "orthopedic",
+  "orthopaedics": "orthopedics",
+  "ostracise": "ostracize",
+  "ostracised": "ostracized",
+  "ostracises": "ostracizes",
+  "ostracising": "ostracizing",
+  "outmanoeuvre": "outmaneuver",
+  "outmanoeuvred": "outmaneuvered",
+  "outmanoeuvres": "outmaneuvers",
+  "outmanoeuvring": "outmaneuvering",
+  "overemphasise": "overemphasize",
+  "overemphasised": "overemphasized",
+  "overemphasises": "overemphasizes",
+  "overemphasising": "overemphasizing",
+  "oxidisation": "oxidization",
+  "oxidise": "oxidize",
+  "oxidised": "oxidized",
+  "oxidises": "oxidizes",
+  "oxidising": "oxidizing",
+  "paederast": "pederast",
+  "paederasts": "pederasts",
+  "paediatric": "pediatric",
+  "paediatrician": "pediatrician",
+  "paediatricians": "pediatricians",
+  "paediatrics": "pediatrics",
+  "paedophile": "pedophile",
+  "paedophiles": "pedophiles",
+  "paedophilia": "pedophilia",
+  "palaeolithic": "paleolithic",
+  "palaeontologist": "paleontologist",
+  "palaeontologists": "paleontologists",
+  "palaeontology": "paleontology",
+  "panelled": "paneled",
+  "panelling": "paneling",
+  "panellist": "panelist",
+  "panellists": "panelists",
+  "paralyse": "paralyze",
+  "paralysed": "paralyzed",
+  "paralyses": "paralyzes",
+  "paralysing": "paralyzing",
+  "parcelled": "parceled",
+  "parcelling": "parceling",
+  "parlour": "parlor",
+  "parlours": "parlors",
+  "particularise": "particularize",
+  "particularised": "particularized",
+  "particularises": "particularizes",
+  "particularising": "particularizing",
+  "passivisation": "passivization",
+  "passivise": "passivize",
+  "passivised": "passivized",
+  "passivises": "passivizes",
+  "passivising": "passivizing",
+  "pasteurisation": "pasteurization",
+  "pasteurise": "pasteurize",
+  "pasteurised": "pasteurized",
+  "pasteurises": "pasteurizes",
+  "pasteurising": "pasteurizing",
+  "patronise": "patronize",
+  "patronised": "patronized",
+  "patronises": "patronizes",
+  "patronising": "patronizing",
+  "patronisingly": "patronizingly",
+  "pedalled": "pedaled",
+  "pedalling": "pedaling",
+  "pedestrianisation": "pedestrianization",
+  "pedestrianise": "pedestrianize",
+  "pedestrianised": "pedestrianized",
+  "pedestrianises": "pedestrianizes",
+  "pedestrianising": "pedestrianizing",
+  "penalise": "penalize",
+  "penalised": "penalized",
+  "penalises": "penalizes",
+  "penalising": "penalizing",
+  "pencilled": "penciled",
+  "pencilling": "penciling",
+  "personalise": "personalize",
+  "personalised": "personalized",
+  "personalises": "personalizes",
+  "personalising": "personalizing",
+  "pharmacopoeia": "pharmacopeia",
+  "pharmacopoeias": "pharmacopeias",
+  "philosophise": "philosophize",
+  "philosophised": "philosophized",
+  "philosophises": "philosophizes",
+  "philosophising": "philosophizing",
+  "philtre": "filter",
+  "philtres": "filters",
+  "phoney": "phony",
+  "plagiarise": "plagiarize",
+  "plagiarised": "plagiarized",
+  "plagiarises": "plagiarizes",
+  "plagiarising": "plagiarizing",
+  "plough": "plow",
+  "ploughed": "plowed",
+  "ploughing": "plowing",
+  "ploughman": "plowman",
+  "ploughmen": "plowmen",
+  "ploughs": "plows",
+  "ploughshare": "plowshare",
+  "ploughshares": "plowshares",
+  "polarisation": "polarization",
+  "polarise": "polarize",
+  "polarised": "polarized",
+  "polarises": "polarizes",
+  "polarising": "polarizing",
+  "politicisation": "politicization",
+  "politicise": "politicize",
+  "politicised": "politicized",
+  "politicises": "politicizes",
+  "politicising": "politicizing",
+  "popularisation": "popularization",
+  "popularise": "popularize",
+  "popularised": "popularized",
+  "popularises": "popularizes",
+  "popularising": "popularizing",
+  "pouffe": "pouf",
+  "pouffes": "poufs",
+  "practise": "practice",
+  "practised": "practiced",
+  "practises": "practices",
+  "practising": "practicing",
+  "praesidium": "presidium",
+  "praesidiums": "presidiums",
+  "pressurisation": "pressurization",
+  "pressurise": "pressurize",
+  "pressurised": "pressurized",
+  "pressurises": "pressurizes",
+  "pressurising": "pressurizing",
+  "pretence": "pretense",
+  "pretences": "pretenses",
+  "primaeval": "primeval",
+  "prioritisation": "prioritization",
+  "prioritise": "prioritize",
+  "prioritised": "prioritized",
+  "prioritises": "prioritizes",
+  "prioritising": "prioritizing",
+  "privatisation": "privatization",
+  "privatisations": "privatizations",
+  "privatise": "privatize",
+  "privatised": "privatized",
+  "privatises": "privatizes",
+  "privatising": "privatizing",
+  "professionalisation": "professionalization",
+  "professionalise": "professionalize",
+  "professionalised": "professionalized",
+  "professionalises": "professionalizes",
+  "professionalising": "professionalizing",
+  "programme": "program",
+  "programmes": "programs",
+  "prologue": "prolog",
+  "prologues": "prologs",
+  "propagandise": "propagandize",
+  "propagandised": "propagandized",
+  "propagandises": "propagandizes",
+  "propagandising": "propagandizing",
+  "proselytise": "proselytize",
+  "proselytised": "proselytized",
+  "proselytiser": "proselytizer",
+  "proselytisers": "proselytizers",
+  "proselytises": "proselytizes",
+  "proselytising": "proselytizing",
+  "psychoanalyse": "psychoanalyze",
+  "psychoanalysed": "psychoanalyzed",
+  "psychoanalyses": "psychoanalyzes",
+  "psychoanalysing": "psychoanalyzing",
+  "publicise": "publicize",
+  "publicised": "publicized",
+  "publicises": "publicizes",
+  "publicising": "publicizing",
+  "pulverisation": "pulverization",
+  "pulverise": "pulverize",
+  "pulverised": "pulverized",
+  "pulverises": "pulverizes",
+  "pulverising": "pulverizing",
+  "pummelled": "pummel",
+  "pummelling": "pummeled",
+  "pyjama": "pajama",
+  "pyjamas": "pajamas",
+  "pzazz": "pizzazz",
+  "quarrelled": "quarreled",
+  "quarrelling": "quarreling",
+  "radicalise": "radicalize",
+  "radicalised": "radicalized",
+  "radicalises": "radicalizes",
+  "radicalising": "radicalizing",
+  "rancour": "rancor",
+  "randomise": "randomize",
+  "randomised": "randomized",
+  "randomises": "randomizes",
+  "randomising": "randomizing",
+  "rationalisation": "rationalization",
+  "rationalisations": "rationalizations",
+  "rationalise": "rationalize",
+  "rationalised": "rationalized",
+  "rationalises": "rationalizes",
+  "rationalising": "rationalizing",
+  "ravelled": "raveled",
+  "ravelling": "raveling",
+  "realisable": "realizable",
+  "realisation": "realization",
+  "realisations": "realizations",
+  "realise": "realize",
+  "realised": "realized",
+  "realises": "realizes",
+  "realising": "realizing",
+  "recognisable": "recognizable",
+  "recognisably": "recognizably",
+  "recognisance": "recognizance",
+  "recognise": "recognize",
+  "recognised": "recognized",
+  "recognises": "recognizes",
+  "recognising": "recognizing",
+  "reconnoitre": "reconnoiter",
+  "reconnoitred": "reconnoitered",
+  "reconnoitres": "reconnoiters",
+  "reconnoitring": "reconnoitering",
+  "refuelled": "refueled",
+  "refuelling": "refueling",
+  "regularisation": "regularization",
+  "regularise": "regularize",
+  "regularised": "regularized",
+  "regularises": "regularizes",
+  "regularising": "regularizing",
+  "remodelled": "remodeled",
+  "remodelling": "remodeling",
+  "remould": "remold",
+  "remoulded": "remolded",
+  "remoulding": "remolding",
+  "remoulds": "remolds",
+  "reorganisation": "reorganization",
+  "reorganisations": "reorganizations",
+  "reorganise": "reorganize",
+  "reorganised": "reorganized",
+  "reorganises": "reorganizes",
+  "reorganising": "reorganizing",
+  "revelled": "reveled",
+  "reveller": "reveler",
+  "revellers": "revelers",
+  "revelling": "reveling",
+  "revitalise": "revitalize",
+  "revitalised": "revitalized",
+  "revitalises": "revitalizes",
+  "revitalising": "revitalizing",
+  "revolutionise": "revolutionize",
+  "revolutionised": "revolutionized",
+  "revolutionises": "revolutionizes",
+  "revolutionising": "revolutionizing",
+  "rhapsodise": "rhapsodize",
+  "rhapsodised": "rhapsodized",
+  "rhapsodises": "rhapsodizes",
+  "rhapsodising": "rhapsodizing",
+  "rigour": "rigor",
+  "rigours": "rigors",
+  "ritualised": "ritualized",
+  "rivalled": "rivaled",
+  "rivalling": "rivaling",
+  "romanticise": "romanticize",
+  "romanticised": "romanticized",
+  "romanticises": "romanticizes",
+  "romanticising": "romanticizing",
+  "rumour": "rumor",
+  "rumoured": "rumored",
+  "rumours": "rumors",
+  "sabre": "saber",
+  "sabres": "sabers",
+  "saltpetre": "saltpeter",
+  "sanitise": "sanitize",
+  "sanitised": "sanitized",
+  "sanitises": "sanitizes",
+  "sanitising": "sanitizing",
+  "satirise": "satirize",
+  "satirised": "satirized",
+  "satirises": "satirizes",
+  "satirising": "satirizing",
+  "saviour": "savior",
+  "saviours": "saviors",
+  "savour": "savor",
+  "savoured": "savored",
+  "savouries": "savories",
+  "savouring": "savoring",
+  "savours": "savors",
+  "savoury": "savory",
+  "scandalise": "scandalize",
+  "scandalised": "scandalized",
+  "scandalises": "scandalizes",
+  "scandalising": "scandalizing",
+  "sceptic": "skeptic",
+  "sceptical": "skeptical",
+  "sceptically": "skeptically",
+  "scepticism": "skepticism",
+  "sceptics": "skeptics",
+  "sceptre": "scepter",
+  "sceptres": "scepters",
+  "scrutinise": "scrutinize",
+  "scrutinised": "scrutinized",
+  "scrutinises": "scrutinizes",
+  "scrutinising": "scrutinizing",
+  "secularisation": "secularization",
+  "secularise": "secularize",
+  "secularised": "secularized",
+  "secularises": "secularizes",
+  "secularising": "secularizing",
+  "sensationalise": "sensationalize",
+  "sensationalised": "sensationalized",
+  "sensationalises": "sensationalizes",
+  "sensationalising": "sensationalizing",
+  "sensitise": "sensitize",
+  "sensitised": "sensitized",
+  "sensitises": "sensitizes",
+  "sensitising": "sensitizing",
+  "sentimentalise": "sentimentalize",
+  "sentimentalised": "sentimentalized",
+  "sentimentalises": "sentimentalizes",
+  "sentimentalising": "sentimentalizing",
+  "sepulchre": "sepulcher",
+  "sepulchres": "sepulchers",
+  "serialisation": "serialization",
+  "serialisations": "serializations",
+  "serialise": "serialize",
+  "serialised": "serialized",
+  "serialises": "serializes",
+  "serialising": "serializing",
+  "sermonise": "sermonize",
+  "sermonised": "sermonized",
+  "sermonises": "sermonizes",
+  "sermonising": "sermonizing",
+  "sheikh": "sheik",
+  "shovelled": "shoveled",
+  "shovelling": "shoveling",
+  "shrivelled": "shriveled",
+  "shrivelling": "shriveling",
+  "signalise": "signalize",
+  "signalised": "signalized",
+  "signalises": "signalizes",
+  "signalising": "signalizing",
+  "signalled": "signaled",
+  "signalling": "signaling",
+  "smoulder": "smolder",
+  "smouldered": "smoldered",
+  "smouldering": "smoldering",
+  "smoulders": "smolders",
+  "snivelled": "sniveled",
+  "snivelling": "sniveling",
+  "snorkelled": "snorkeled",
+  "snorkelling": "snorkeling",
+  "snowplough": "snowplow",
+  "snowploughs": "snowplow",
+  "socialisation": "socialization",
+  "socialise": "socialize",
+  "socialised": "socialized",
+  "socialises": "socializes",
+  "socialising": "socializing",
+  "sodomise": "sodomize",
+  "sodomised": "sodomized",
+  "sodomises": "sodomizes",
+  "sodomising": "sodomizing",
+  "solemnise": "solemnize",
+  "solemnised": "solemnized",
+  "solemnises": "solemnizes",
+  "solemnising": "solemnizing",
+  "sombre": "somber",
+  "specialisation": "specialization",
+  "specialisations": "specializations",
+  "specialise": "specialize",
+  "specialised": "specialized",
+  "specialises": "specializes",
+  "specialising": "specializing",
+  "spectre": "specter",
+  "spectres": "specters",
+  "spiralled": "spiraled",
+  "spiralling": "spiraling",
+  "splendour": "splendor",
+  "splendours": "splendors",
+  "squirrelled": "squirreled",
+  "squirrelling": "squirreling",
+  "stabilisation": "stabilization",
+  "stabilise": "stabilize",
+  "stabilised": "stabilized",
+  "stabiliser": "stabilizer",
+  "stabilisers": "stabilizers",
+  "stabilises": "stabilizes",
+  "stabilising": "stabilizing",
+  "standardisation": "standardization",
+  "standardise": "standardize",
+  "standardised": "standardized",
+  "standardises": "standardizes",
+  "standardising": "standardizing",
+  "stencilled": "stenciled",
+  "stencilling": "stenciling",
+  "sterilisation": "sterilization",
+  "sterilisations": "sterilizations",
+  "sterilise": "sterilize",
+  "sterilised": "sterilized",
+  "steriliser": "sterilizer",
+  "sterilisers": "sterilizers",
+  "sterilises": "sterilizes",
+  "sterilising": "sterilizing",
+  "stigmatisation": "stigmatization",
+  "stigmatise": "stigmatize",
+  "stigmatised": "stigmatized",
+  "stigmatises": "stigmatizes",
+  "stigmatising": "stigmatizing",
+  "storey": "story",
+  "storeys": "stories",
+  "subsidisation": "subsidization",
+  "subsidise": "subsidize",
+  "subsidised": "subsidized",
+  "subsidiser": "subsidizer",
+  "subsidisers": "subsidizers",
+  "subsidises": "subsidizes",
+  "subsidising": "subsidizing",
+  "succour": "succor",
+  "succoured": "succored",
+  "succouring": "succoring",
+  "succours": "succors",
+  "sulphate": "sulfate",
+  "sulphates": "sulfates",
+  "sulphide": "sulfide",
+  "sulphides": "sulfides",
+  "sulphur": "sulfur",
+  "sulphurous": "sulfurous",
+  "summarise": "summarize",
+  "summarised": "summarized",
+  "summarises": "summarizes",
+  "summarising": "summarizing",
+  "swivelled": "swiveled",
+  "swivelling": "swiveling",
+  "symbolise": "symbolize",
+  "symbolised": "symbolized",
+  "symbolises": "symbolizes",
+  "symbolising": "symbolizing",
+  "sympathise": "sympathize",
+  "sympathised": "sympathized",
+  "sympathiser": "sympathizer",
+  "sympathisers": "sympathizers",
+  "sympathises": "sympathizes",
+  "sympathising": "sympathizing",
+  "synchronisation": "synchronization",
+  "synchronise": "synchronize",
+  "synchronised": "synchronized",
+  "synchronises": "synchronizes",
+  "synchronising": "synchronizing",
+  "synthesise": "synthesize",
+  "synthesised": "synthesized",
+  "synthesiser": "synthesizer",
+  "synthesisers": "synthesizers",
+  "synthesises": "synthesizes",
+  "synthesising": "synthesizing",
+  "syphon": "siphon",
+  "syphoned": "siphoned",
+  "syphoning": "siphoning",
+  "syphons": "siphons",
+  "systematisation": "systematization",
+  "systematise": "systematize",
+  "systematised": "systematized",
+  "systematises": "systematizes",
+  "systematising": "systematizing",
+  "tantalise": "tantalize",
+  "tantalised": "tantalized",
+  "tantalises": "tantalizes",
+  "tantalising": "tantalizing",
+  "tantalisingly": "tantalizingly",
+  "tasselled": "tasseled",
+  "technicolour": "technicolor",
+  "temporise": "temporize",
+  "temporised": "temporized",
+  "temporises": "temporizes",
+  "temporising": "temporizing",
+  "tenderise": "tenderize",
+  "tenderised": "tenderized",
+  "tenderises": "tenderizes",
+  "tenderising": "tenderizing",
+  "terrorise": "terrorize",
+  "terrorised": "terrorized",
+  "terrorises": "terrorizes",
+  "terrorising": "terrorizing",
+  "theatre": "theater",
+  "theatregoer": "theatergoer",
+  "theatregoers": "theatergoers",
+  "theatres": "theaters",
+  "theorise": "theorize",
+  "theorised": "theorized",
+  "theorises": "theorizes",
+  "theorising": "theorizing",
+  "tonne": "ton",
+  "tonnes": "tons",
+  "towelled": "toweled",
+  "towelling": "toweling",
+  "toxaemia": "toxemia",
+  "tranquillise": "tranquilize",
+  "tranquillised": "tranquilized",
+  "tranquilliser": "tranquilizer",
+  "tranquillisers": "tranquilizers",
+  "tranquillises": "tranquilizes",
+  "tranquillising": "tranquilizing",
+  "tranquillity": "tranquility",
+  "tranquillize": "tranquilize",
+  "tranquillized": "tranquilized",
+  "tranquillizer": "tranquilizer",
+  "tranquillizers": "tranquilizers",
+  "tranquillizes": "tranquilizes",
+  "tranquillizing": "tranquilizing",
+  "tranquilly": "tranquility",
+  "transistorised": "transistorized",
+  "traumatise": "traumatize",
+  "traumatised": "traumatized",
+  "traumatises": "traumatizes",
+  "traumatising": "traumatizing",
+  "travelled": "traveled",
+  "traveller": "traveler",
+  "travellers": "travelers",
+  "travelling": "traveling",
+  "travelog": "travelogue",
+  "travelogs": "travelogues",
+  "trialled": "trialed",
+  "trialling": "trialing",
+  "tricolour": "tricolor",
+  "tricolours": "tricolors",
+  "trivialise": "trivialize",
+  "trivialised": "trivialized",
+  "trivialises": "trivializes",
+  "trivialising": "trivializing",
+  "tumour": "tumor",
+  "tumours": "tumors",
+  "tunnelled": "tunneled",
+  "tunnelling": "tunneling",
+  "tyrannise": "tyrannize",
+  "tyrannised": "tyrannized",
+  "tyrannises": "tyrannizes",
+  "tyrannising": "tyrannizing",
+  "tyre": "tire",
+  "tyres": "tires",
+  "unauthorised": "unauthorized",
+  "uncivilised": "uncivilized",
+  "underutilised": "underutilized",
+  "unequalled": "unequaled",
+  "unfavourable": "unfavorable",
+  "unfavourably": "unfavorably",
+  "unionisation": "unionization",
+  "unionise": "unionize",
+  "unionised": "unionized",
+  "unionises": "unionizes",
+  "unionising": "unionizing",
+  "unorganised": "unorganized",
+  "unravelled": "unraveled",
+  "unravelling": "unraveling",
+  "unrecognisable": "unrecognizable",
+  "unrecognised": "unrecognized",
+  "unrivalled": "unrivaled",
+  "unsavoury": "unsavory",
+  "untrammelled": "untrammeled",
+  "urbanisation": "urbanization",
+  "urbanise": "urbanize",
+  "urbanised": "urbanized",
+  "urbanises": "urbanizes",
+  "urbanising": "urbanizing",
+  "utilisable": "utilizable",
+  "utilisation": "utilization",
+  "utilise": "utilize",
+  "utilised": "utilized",
+  "utilises": "utilizes",
+  "utilising": "utilizing",
+  "valour": "valor",
+  "vandalise": "vandalize",
+  "vandalised": "vandalized",
+  "vandalises": "vandalizes",
+  "vandalising": "vandalizing",
+  "vaporisation": "vaporization",
+  "vaporise": "vaporize",
+  "vaporised": "vaporized",
+  "vaporises": "vaporizes",
+  "vaporising": "vaporizing",
+  "vapour": "vapor",
+  "vapours": "vapors",
+  "verbalise": "verbalize",
+  "verbalised": "verbalized",
+  "verbalises": "verbalizes",
+  "verbalising": "verbalizing",
+  "victimisation": "victimization",
+  "victimise": "victimize",
+  "victimised": "victimized",
+  "victimises": "victimizes",
+  "victimising": "victimizing",
+  "videodisc": "videodisk",
+  "videodiscs": "videodisks",
+  "vigour": "vigor",
+  "visualisation": "visualization",
+  "visualisations": "visualizations",
+  "visualise": "visualize",
+  "visualised": "visualized",
+  "visualises": "visualizes",
+  "visualising": "visualizing",
+  "vocalisation": "vocalization",
+  "vocalisations": "vocalizations",
+  "vocalise": "vocalize",
+  "vocalised": "vocalized",
+  "vocalises": "vocalizes",
+  "vocalising": "vocalizing",
+  "vulcanised": "vulcanized",
+  "vulgarisation": "vulgarization",
+  "vulgarise": "vulgarize",
+  "vulgarised": "vulgarized",
+  "vulgarises": "vulgarizes",
+  "vulgarising": "vulgarizing",
+  "waggon": "wagon",
+  "waggons": "wagons",
+  "watercolour": "watercolor",
+  "watercolours": "watercolors",
+  "weaselled": "weaseled",
+  "weaselling": "weaseling",
+  "westernisation": "westernization",
+  "westernise": "westernize",
+  "westernised": "westernized",
+  "westernises": "westernizes",
+  "westernising": "westernizing",
+  "womanise": "womanize",
+  "womanised": "womanized",
+  "womaniser": "womanizer",
+  "womanisers": "womanizers",
+  "womanises": "womanizes",
+  "womanising": "womanizing",
+  "woollen": "woolen",
+  "woollens": "woolens",
+  "woollies": "woolies",
+  "woolly": "wooly",
+  "worshipped": "worshiped",
+  "worshipper": "worshiper",
+  "worshipping": "worshiping",
+  "yodelled": "yodeled",
+  "yodelling": "yodeling",
+  "yoghourt": "yogurt",
+  "yoghourts": "yogurts",
+  "yoghurt": "yogurt",
+  "yoghurts": "yogurts"
+}

patch_func.py

@@ -0,0 +1,267 @@
+import torch
+import torch.nn as nn
+
+
+def partial_rope_mask(model_args, mha2mla_args):
+    """
+    Generate different types of masks for partial rotary position embeddings (RoPE)
+    based on configuration settings.
+    Returns:
+        Appropriate mask tensor based on the specified version
+    """
+    n_head = model_args.encoder_attention_heads
+    # MODIFIED (MHA adaptation): upstream hardcoded n_k_head=1 (a GQA / MLA-target assumption).
+    # Whisper is MHA — its k has n_head heads — so the k-side rope mask must span all heads for
+    # partial-preservation to partition keys correctly. (No effect on full_k_svd, which ignores
+    # k_c_indices and deletes k_r_proj.)
+    n_k_head = n_head
+    if hasattr(model_args, "head_dim") and model_args.head_dim:
+        d_head = model_args.head_dim
+    else:
+        # MODIFIED (prior-art §6 fix #1): upstream used d_head = d_model here (768 for
+        # whisper-small), corrupting the rope mask; the per-head dim is d_model // n_head.
+        d_head = model_args.d_model // n_head
+    d_head_half = d_head // 2
+    rope_dim_for_mla = mha2mla_args.rope_dim_for_mla
+    rope_dim_for_mla_half = rope_dim_for_mla // 2
+    rope_version = mha2mla_args.partial_rope_version
+    mask = torch.zeros(d_head)
+
+    def select_high_frequency(mask):
+        """
+        Select high-frequency components (first rope_dim_for_mla dimensions)
+        Returns:
+            mask: Binary mask with 1s for the first rope_dim_for_mla dimensions
+        """
+        mask[:rope_dim_for_mla_half] = 1
+        mask[d_head_half : d_head_half + rope_dim_for_mla_half] = 1
+        q_masks = mask.repeat(n_head).bool()
+        k_masks = mask.repeat(n_k_head).bool()
+        return q_masks, k_masks
+
+    def select_low_frequency(mask):
+        """
+        Select low-frequency components (last rope_dim_for_mla dimensions)
+        Returns:
+            mask: Binary mask with 1s for the last rope_dim_for_mla dimensions
+        """
+        mask[d_head - rope_dim_for_mla_half :] = 1
+        mask[d_head_half - rope_dim_for_mla_half : d_head_half] = 1
+        q_masks = mask.repeat(n_head).bool()
+        k_masks = mask.repeat(n_k_head).bool()
+        return q_masks, k_masks
+
+    def select_uniform_frequency(mask, start_point):
+        """
+        Select uniformly distributed dimensions for RoPE
+        Returns:
+            mask: Binary mask with 1s at uniformly spaced positions
+        """
+        step = d_head // rope_dim_for_mla
+        assert d_head_half % step == 0, "rope_dim_for_mla must be greater than 0"
+
+        for i in range(start_point, d_head, step):
+            mask[i] = 1
+        # MODIFIED (prior-art §6): repeat across heads like the high/low paths — q over n_head,
+        # k over n_k_head. Upstream returned the un-repeated per-head mask, which mis-sizes q_proj
+        # for MHA Whisper once d_head is the correct per-head 64 (fix #1).
+        q_masks = mask.repeat(n_head).bool()
+        k_masks = mask.repeat(n_k_head).bool()
+        return q_masks, k_masks
+
+    def select_2norm_frequency(mask, rope_dim_for_mla):
+        """
+        Select dimensions based on 2-norm frequency importance
+        Returns:
+            mask: Binary mask with 1s for the top rope_dim_for_mla dimensions by 2-norm
+        """
+        # This is a placeholder implementation since the exact 2-norm selection
+        # method was not detailed in the comments. In practice, this would
+        # require statistics from the weight matrices to determine importance.
+        with open(mha2mla_args.qk_tensor_path, "rb") as fin:
+            qk_norm_rank = torch.load(fin, weights_only=True)
+
+        k_masks = qk_norm_rank < rope_dim_for_mla_half
+        if mha2mla_args.is_gqa2mha2mla:
+            q_masks = k_masks
+        else:
+            q_masks = k_masks
+        k_masks = k_masks.view(k_masks.size(0), -1)
+        q_masks = q_masks.view(q_masks.size(0), -1)
+        return q_masks, k_masks
+
+    if rope_version == "high":
+        return select_high_frequency(mask)
+    elif rope_version == "low":
+        return select_low_frequency(mask)
+    elif rope_version == "uniform":
+        return select_uniform_frequency(mask, mha2mla_args.uniform_start_point)
+    elif rope_version == "2-norm":
+        return select_2norm_frequency(d_head, rope_dim_for_mla)
+
+
+class LowRankKVLinear(nn.Module):
+    """
+    A low-rank approximation of a linear layer.
+    Instead of storing a full matrix W of shape (out_features, in_features),
+    it stores two matrices: down of shape (in_features, low_rank) and
+    up of shape (low_rank, out_features), such that W ≈ up.T @ down.T
+    """
+
+    def __init__(
+        self,
+        d_in,
+        d_k_out,
+        d_v_out,
+        d_mid=0,
+        k_approx=False,
+        v_approx=False,
+        kv_joint=False,
+        bias=None,
+    ):
+        super().__init__()
+        # TODO: add activations after down_kv
+        if kv_joint:  # (x * W_q）* （x * down_kv * up_k）
+            self.down_kv = nn.Linear(in_features=d_in, out_features=d_mid, bias=False)
+        if not kv_joint and k_approx:
+            self.down_k = nn.Linear(in_features=d_in, out_features=d_mid, bias=False)
+        if not kv_joint and v_approx:
+            self.down_v = nn.Linear(in_features=d_in, out_features=d_mid, bias=False)
+
+        d_k_in = d_mid if k_approx else d_in
+        self.up_k = nn.Linear(in_features=d_k_in, out_features=d_k_out, bias=True)
+        d_v_in = d_mid if v_approx else d_in
+        self.up_v = nn.Linear(in_features=d_v_in, out_features=d_v_out, bias=True)
+
+    def reset_parameters(
+        self,
+        down_kv_weight=None,
+        down_k_weight=None,
+        down_v_weight=None,
+        up_k_weight=None,
+        up_v_weight=None,
+        up_k_bias=None,
+        up_v_bias=None,
+    ):
+        if down_kv_weight is not None:
+            self.down_kv.weight.data.copy_(down_kv_weight)
+        if down_k_weight is not None:
+            self.down_k.weight.data.copy_(down_k_weight)
+        if down_v_weight is not None:
+            self.down_v.weight.data.copy_(down_v_weight)
+        if up_k_weight is not None:
+            self.up_k.weight.data.copy_(up_k_weight)
+        if up_v_weight is not None:
+            self.up_v.weight.data.copy_(up_v_weight)
+        if up_k_bias is not None:
+            self.up_k.bias.data.copy_(up_k_bias)
+        if up_v_bias is not None:
+            self.up_v.bias.data.copy_(up_v_bias)
+
+    def mha_forward(self, x):
+        # x: (batch_size, seq_len, in_features)
+        if hasattr(self, "down_kv"):
+            k = v = self.down_kv(x)
+        else:
+            k = self.down_k(x) if hasattr(self, "down_k") else x
+            v = self.down_v(x) if hasattr(self, "down_v") else x
+        k = self.up_k(k)
+        v = self.up_v(v)
+        return k, v
+
+    def kv_forward(self, x):
+        kv = self.down_kv(x)
+        # TODO: Triton kernel
+        return kv
+
+    def kv_up(self, x):
+        k = self.up_k(x)
+        v = self.up_v(x)
+        return k, v
+
+
+def SVD(X, r):
+    U, S, V = torch.linalg.svd(X.to(torch.float32), full_matrices=False)
+    U, S, V = U[:, :r], S[:r], V[:r, :]
+    U @= torch.diag(S)
+    return V, U
+
+
+def svd_low_rank_approx(k_c_weight, k_c_bias, v_weight, v_bias, d_kv_mid, method):
+    d_k_c, d_v, d_kv_in = k_c_weight.size(0), v_weight.size(0), v_weight.size(1)
+    has_bias = k_c_bias is not None
+
+    if method == "only_key":
+        down_k, up_k = SVD(k_c_weight, d_kv_mid)
+        kv_proj = LowRankKVLinear(
+            d_kv_in, d_k_c, d_v, d_mid=d_kv_mid, k_approx=True, bias=has_bias
+        )
+        kv_proj.reset_parameters(
+            down_k_weight=down_k,
+            up_k_weight=up_k,
+            up_v_weight=v_weight,
+            up_k_bias=k_c_bias,
+            up_v_bias=v_bias,
+        )
+    elif method == "only_value":
+        down_v, up_v = SVD(v_weight, d_kv_mid)
+        kv_proj = LowRankKVLinear(
+            d_kv_in, d_k_c, d_v, d_mid=d_kv_mid, v_approx=True, bias=has_bias
+        )
+        kv_proj.reset_parameters(
+            down_v_weight=down_v,
+            up_k_weight=k_c_weight,
+            up_v_weight=up_v,
+            up_k_bias=k_c_bias,
+            up_v_bias=v_bias,
+        )
+    elif method == "split":
+        down_k, up_k = SVD(k_c_weight, d_kv_mid)
+        down_v, up_v = SVD(v_weight, d_kv_mid)
+        kv_proj = LowRankKVLinear(
+            d_kv_in,
+            d_k_c,
+            d_v,
+            d_mid=d_kv_mid,
+            k_approx=True,
+            v_approx=True,
+            bias=has_bias,
+        )
+        kv_proj.reset_parameters(
+            down_k_weight=down_k,
+            down_v_weight=down_v,
+            up_k_weight=up_k,
+            up_v_weight=up_v,
+            up_k_bias=k_c_bias,
+            up_v_bias=v_bias,
+        )
+    elif method == "joint":
+        joint_kv = torch.cat([k_c_weight, v_weight])
+        down_kv, up_kv = SVD(joint_kv, d_kv_mid)
+        up_k, up_v = up_kv.split([d_k_c, d_v])
+        kv_proj = LowRankKVLinear(
+            d_kv_in,
+            d_k_c,
+            d_v,
+            d_mid=d_kv_mid,
+            k_approx=True,
+            v_approx=True,
+            kv_joint=True,
+            bias=has_bias,
+        )
+        kv_proj.reset_parameters(
+            down_kv_weight=down_kv,
+            up_k_weight=up_k,
+            up_v_weight=up_v,
+            up_k_bias=k_c_bias,
+            up_v_bias=v_bias,
+        )
+    elif method == "none":
+        kv_proj = LowRankKVLinear(d_kv_in, d_k_c, d_v, bias=has_bias)
+        kv_proj.reset_parameters(
+            up_k_weight=k_c_weight,
+            up_v_weight=v_weight,
+            up_k_bias=k_c_bias,
+            up_v_bias=v_bias,
+        )
+    return kv_proj

patching_model_load.py

@@ -0,0 +1,159 @@
+import logging
+
+import torch
+import torch.nn as nn
+
+logger = logging.getLogger(__name__)
+
+from .patch_func import partial_rope_mask, svd_low_rank_approx
+from . import modeling_whisper
+from .modeling_whisper import WhisperAttention
+
+
+def reorder_matrix_rows(mask, is_cat=False):
+    """
+    Reorder rows in a matrix based on a binary mask.
+    Rows corresponding to 1s in the mask come first, then rows corresponding to 0s.
+
+    Args:
+        weight: The weight matrix to reorder
+        mask: A binary mask (list or tensor) of length equal to weight.shape[0]
+
+    Returns:
+        The reordered weight matrix
+    """
+    ones_indices = torch.where(mask)[0]
+    zeros_indices = torch.where(~mask)[0]
+    if is_cat:
+        return torch.cat([ones_indices, zeros_indices])
+    else:
+        return ones_indices, zeros_indices
+
+
+def patch_model(model, model_args, mha2mla_args):
+    """
+    Patch a huggingface model by:
+    1. Reordering rows in q_proj and k_proj matrices based on a partial-rope mask
+    2. Replacing v_proj with a low-rank approximation
+
+    Args:
+        model: The Qwen2ForCausalLM model
+        mask: A binary mask for reordering q_proj and k_proj
+        low_rank: The rank for the low-rank approximation of v_proj
+    """
+    q_masks, k_masks = partial_rope_mask(model_args, mha2mla_args)
+
+    n_head = model_args.encoder_attention_heads
+    n_k_head = n_head  # MHA: k has n_head heads (matches partial_rope_mask)
+    if hasattr(model_args, "head_dim"):
+        d_head = model_args.head_dim
+    else:
+        d_head = model_args.d_model // n_head
+    q_idx = []
+    k_idx = []
+    layer_idx = 0
+    targets = getattr(mha2mla_args, "targets", None)   # set of roles to convert; None = all (§6 #2)
+    preserve = getattr(mha2mla_args, "preserve_rope", False)
+    for name, layer in model.named_modules():
+        if not isinstance(layer, WhisperAttention):
+            continue
+        # scope gating: classify the module by name and skip if outside the requested scope.
+        if "encoder_attn" in name:
+            role = "decoder_cross"
+        elif "decoder" in name:
+            role = "decoder_self"
+        else:
+            role = "encoder_self"
+        if targets is not None and role not in targets:
+            continue
+        q_weight = layer.q_proj.weight
+        q_bias = getattr(layer.q_proj, "bias", None)
+
+        # Reorder q_proj rows (rope dims first) ONLY for partial_preservation, whose split forward
+        # regroups per-head. full_k_svd keeps k in original order (nosplit forward), so permuting q
+        # there would twist q.k — leave q untouched.
+        q_mask = q_masks[layer_idx] if len(q_masks.shape) == 2 else q_masks
+        q_indices = reorder_matrix_rows(q_mask, is_cat=True)
+        if preserve:
+            layer.q_proj.weight.data.copy_(q_weight[q_indices])
+            if q_bias is not None:
+                layer.q_proj.bias.data.copy_(q_bias[q_indices])
+
+        # 2. Reorder k_proj and setup k_r_proj
+        # Get original weights and biases if biases exist
+        k_weight = layer.k_proj.weight
+        if mha2mla_args.is_gqa2mha2mla:
+            k_weight = (
+                k_weight.view(n_k_head, -1, k_weight.size(-1))
+                .repeat_interleave(n_head // n_k_head, dim=0)
+                .view(-1, k_weight.size(-1))
+            )
+        k_bias = getattr(layer.k_proj, "bias", None)
+
+        # Reorder and update weights and biases if biases exist
+        k_mask = k_masks[layer_idx] if len(k_masks.shape) == 2 else k_masks
+        k_r_indices, k_c_indices = reorder_matrix_rows(k_mask, is_cat=False)
+        k_r_proj = nn.Linear(k_weight.size(1), k_r_indices.size(0), k_bias is not None)
+        k_r_proj.weight.data.copy_(k_weight[k_r_indices])
+        if k_bias is not None:
+            k_r_proj.bias.data.copy_(k_bias[k_r_indices])
+        layer.k_r_proj = k_r_proj
+
+        # Reorder Q_Norm/K_Norm (e.g., Qwen3) if exist
+        if hasattr(layer, "q_norm"):
+            # TODO: support olmo's Q_Norm/K_Norm
+            assert mha2mla_args.partial_rope_version != "2_norm", (
+                "Qwen3 does not suppert 2_norm yet."
+            )
+            norm_indices = reorder_matrix_rows(q_mask[:d_head], is_cat=True)
+            qn_weight = layer.self_attn.q_norm.weight
+            kn_weight = layer.self_attn.k_norm.weight
+            layer.self_attn.q_norm.weight.data.copy_(qn_weight[norm_indices])
+            layer.self_attn.k_norm.weight.data.copy_(kn_weight[norm_indices])
+
+        # 3. Setup low-rank kv_proj (§6 #3): partial-preservation compresses only the non-rope
+        #    rows k_weight[k_c_indices]; full_k_svd compresses the full key.
+        if preserve:
+            k_c_weight = k_weight[k_c_indices]
+            k_c_bias = k_bias[k_c_indices] if k_bias is not None else None
+        else:
+            k_c_weight = k_weight
+            k_c_bias = k_bias if k_bias is not None else None
+        kv_proj = svd_low_rank_approx(
+            k_c_weight=k_c_weight,
+            k_c_bias=k_c_bias,
+            v_weight=layer.v_proj.weight,
+            v_bias=getattr(layer.v_proj, "bias", None),
+            d_kv_mid=mha2mla_args.low_rank * model_args.num_key_value_heads,
+            method=mha2mla_args.svd_init_method,
+        )
+        layer.kv_proj = kv_proj
+
+        # 4. Delete original k_proj/v_proj; keep k_r_proj only for partial-preservation (§6 #4 —
+        #    the split forward reconstructs key = cat(k_r_proj(x), kv_proj.up_k(...))).
+        delattr(layer, "k_proj")
+        delattr(layer, "v_proj")
+        if not preserve:
+            delattr(layer, "k_r_proj")
+
+        d_q_r = n_head * mha2mla_args.rope_dim_for_mla
+        q_idx.append(q_indices[:d_q_r])
+        k_idx.append(k_r_indices)
+        logger.debug("Layer %d: set up q_proj, k_r_proj, and kv_proj", layer_idx)
+
+        # 5. [Optional] Randomly init if is_mla_from_scratch == True
+        if mha2mla_args.is_mla_from_scratch:
+            std = model_args.initializer_range
+            for name, param in layer.self_attn.named_parameters():
+                if "o_proj" in name or "proj" not in name:
+                    continue
+                if "weight" in name:
+                    param.data.normal_(mean=0.0, std=std)
+                    print(f"Reinit {name}")
+                if "bias" in name:
+                    param.data.zero_()
+                    print(f"Reinit {name}")
+
+        layer_idx = layer_idx + 1
+
+    return model, q_idx, k_idx

patching_whisper.py

@@ -0,0 +1,281 @@
+import inspect
+from typing import Callable, Optional, Tuple
+
+import torch
+from transformers import Cache
+from transformers.utils import logging
+from transformers.processing_utils import Unpack
+from . import modeling_whisper
+from .modeling_whisper import WhisperAttention, WhisperDecoderLayer
+from torch import nn
+import types
+
+
+logger = logging.get_logger(__name__)
+
+
+def _mla_split_heads(states, num_heads, n_rope):
+    """Reshape a [rope(all heads) | pass(all heads)] projection to (bsz, num_heads, seq, head_dim)
+    WITHOUT scrambling heads: split the rope/pass sections and reshape each PER-HEAD, then concat
+    within head. Fixes the cross-head reshape flaw (a flat reshape pooled dims from different
+    original heads under one softmax)."""
+    bsz, seq, dim = states.shape
+    rope = states[..., :n_rope].reshape(bsz, seq, num_heads, n_rope // num_heads)
+    pas = states[..., n_rope:].reshape(bsz, seq, num_heads, (dim - n_rope) // num_heads)
+    return torch.cat([rope, pas], dim=-1).transpose(1, 2).contiguous()
+
+
+def _mla_key_from_parts(key_r, key_c, num_heads):
+    """Assemble the split-scheme key per-head from its preserved (key_r) + compressed (key_c) parts,
+    keeping heads intact -> (bsz, num_heads, seq, head_dim)."""
+    bsz, seq = key_r.shape[0], key_r.shape[1]
+    r = key_r.reshape(bsz, seq, num_heads, key_r.size(-1) // num_heads)
+    c = key_c.reshape(bsz, seq, num_heads, key_c.size(-1) // num_heads)
+    return torch.cat([r, c], dim=-1).transpose(1, 2).contiguous()
+
+
+def custom_WhisperAttention_mla_forward(
+    self,
+    hidden_states: torch.Tensor,
+    attention_mask: Optional[torch.Tensor],
+    past_key_value: Optional[Cache] = None,
+    key_value_states: Optional[torch.Tensor] = None,
+    cache_position: Optional[torch.LongTensor] = None,
+    layer_head_mask: Optional[torch.Tensor] = None,
+    output_attentions: bool = False,
+) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+    is_cross_attention = key_value_states is not None
+    bsz, tgt_len, _ = hidden_states.size()
+    input_shape = hidden_states.shape[:-1]
+    kv_shape = (*input_shape, 1, -1)
+    num_q_heads = 12
+    q_shape = (*input_shape, num_q_heads, -1)
+    query_states = self.q_proj(hidden_states) * self.scaling
+    if (
+        is_cross_attention
+        and past_key_value is not None
+        and past_key_value[0].shape[1] == key_value_states.shape[1]
+    ):
+        # reuse k,v, cross_attentions
+        kv_states = past_key_value[0]
+        key_r_states = past_key_value[1]
+        key_c_states, value_states = self.kv_proj.kv_up(kv_states)
+        key_states = _mla_key_from_parts(key_r_states, key_c_states, self.num_heads)
+        value_states = self._shape(value_states, -1, bsz)
+    # NOTE: value_states = self.v_proj(hidden_states)
+    elif is_cross_attention:
+        kv_states = self.kv_proj.kv_forward(key_value_states)
+        key_c_states, value_states = self.kv_proj.kv_up(kv_states)
+        key_r_states = self.k_r_proj(key_value_states)
+        key_states = _mla_key_from_parts(key_r_states, key_c_states, self.num_heads)
+        value_states = self._shape(value_states, -1, bsz)
+    elif past_key_value is not None:
+        kv_states = self.kv_proj.kv_forward(hidden_states)
+        kv_states = torch.cat([past_key_value[0], kv_states], dim=1)
+        key_c_states, value_states = self.kv_proj.kv_up(kv_states)
+        key_r_states = self.k_r_proj(hidden_states)
+        key_r_states = torch.cat([past_key_value[1], key_r_states], dim=1)
+        key_states = _mla_key_from_parts(key_r_states, key_c_states, self.num_heads)
+        value_states = self._shape(value_states, -1, bsz)
+    else:
+        kv_states = self.kv_proj.kv_forward(hidden_states)
+        key_c_states, value_states = self.kv_proj.kv_up(kv_states)
+        key_r_states = self.k_r_proj(hidden_states)
+        key_states = _mla_key_from_parts(key_r_states, key_c_states, self.num_heads)
+        value_states = self._shape(value_states, -1, bsz)
+        
+    # NOTE: the code below has not been modified.
+    if self.is_decoder:
+            # if cross_attention save Tuple(torch.Tensor, torch.Tensor) of all cross attention key/value_states.
+            # Further calls to cross_attention layer can then reuse all cross-attention
+            # key/value_states (first "if" case)
+            # if uni-directional self-attention (decoder) save Tuple(torch.Tensor, torch.Tensor) of
+            # all previous decoder key/value_states. Further calls to uni-directional self-attention
+            # can concat previous decoder key/value_states to current projected key/value_states (third "elif" case)
+            # if encoder bi-directional self-attention `past_key_value` is always `None`
+        past_key_value = (kv_states, key_r_states)
+
+    proj_shape = (bsz * self.num_heads, -1, self.head_dim)
+    # PER-HEAD reshape (q is reordered [rope|pass]); rope section width == key_r width.
+    query_states = _mla_split_heads(query_states, self.num_heads, key_r_states.size(-1)).reshape(*proj_shape)
+    key_states = key_states.reshape(*proj_shape)
+    value_states = value_states.reshape(*proj_shape)
+
+    src_len = key_states.size(1)
+    attn_weights = torch.bmm(query_states, key_states.transpose(1, 2))
+
+    if attn_weights.size() != (bsz * self.num_heads, tgt_len, src_len):
+        raise ValueError(
+            f"Attention weights should be of size {(bsz * self.num_heads, tgt_len, src_len)}, but is"
+            f" {attn_weights.size()}"
+        )
+
+    if attention_mask is not None:
+        if attention_mask.size() != (bsz, 1, tgt_len, src_len):
+            raise ValueError(
+                f"Attention mask should be of size {(bsz, 1, tgt_len, src_len)}, but is {attention_mask.size()}"
+                )
+        attn_weights = attn_weights.view(bsz, self.num_heads, tgt_len, src_len) + attention_mask
+        attn_weights = attn_weights.view(bsz * self.num_heads, tgt_len, src_len)
+
+    attn_weights = nn.functional.softmax(attn_weights, dim=-1)
+
+    if layer_head_mask is not None:
+        if layer_head_mask.size() != (self.num_heads,):
+            raise ValueError(
+                f"Head mask for a single layer should be of size {(self.num_heads,)}, but is"
+                f" {layer_head_mask.size()}"
+            )
+        attn_weights = layer_head_mask.view(1, -1, 1, 1) * attn_weights.view(bsz, self.num_heads, tgt_len, src_len)
+        attn_weights = attn_weights.view(bsz * self.num_heads, tgt_len, src_len)
+
+    if output_attentions:
+        attn_weights_reshaped = attn_weights.view(bsz, self.num_heads, tgt_len, src_len)
+        attn_weights = attn_weights_reshaped.view(bsz * self.num_heads, tgt_len, src_len)
+    else:
+        attn_weights_reshaped = None
+
+    attn_probs = nn.functional.dropout(attn_weights, p=self.dropout, training=self.training)
+
+    attn_output = torch.bmm(attn_probs, value_states)
+
+    if attn_output.size() != (bsz * self.num_heads, tgt_len, self.head_dim):
+        raise ValueError(
+            f"`attn_output` should be of size {(bsz * self.num_heads, tgt_len, self.head_dim)}, but is"
+            f" {attn_output.size()}"
+        )
+
+    attn_output = attn_output.view(bsz, self.num_heads, tgt_len, self.head_dim)
+    attn_output = attn_output.transpose(1, 2)
+
+    attn_output = attn_output.reshape(bsz, tgt_len, self.embed_dim)
+
+    attn_output = self.out_proj(attn_output)
+   
+    return attn_output, attn_weights_reshaped, past_key_value
+
+
+def custom_WhisperAttention_nosplit_forward(
+    self,
+    hidden_states: torch.Tensor,
+    attention_mask: Optional[torch.Tensor],
+    past_key_value: Optional[Cache] = None,
+    key_value_states: Optional[torch.Tensor] = None,
+    cache_position: Optional[torch.LongTensor] = None,
+    layer_head_mask: Optional[torch.Tensor] = None,
+    output_attentions: bool = False,
+) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+    is_cross_attention = key_value_states is not None
+    bsz, tgt_len, _ = hidden_states.size()
+    input_shape = hidden_states.shape[:-1]
+    kv_shape = (*input_shape, 1, -1)
+    num_q_heads = 12
+    q_shape = (*input_shape, num_q_heads, -1)
+    query_states = self.q_proj(hidden_states) * self.scaling
+    if (
+        is_cross_attention
+        and past_key_value is not None
+        and past_key_value[0].shape[1] == key_value_states.shape[1]
+    ):
+        # reuse k,v, cross_attentions
+        kv_states = past_key_value[0]
+        key_states, value_states = self.kv_proj.kv_up(kv_states)
+        key_states = self._shape(key_states, -1, bsz)
+        value_states = self._shape(value_states, -1, bsz)
+    # NOTE: value_states = self.v_proj(hidden_states)
+    elif is_cross_attention:
+        kv_states = self.kv_proj.kv_forward(key_value_states)
+        key_states, value_states = self.kv_proj.kv_up(kv_states)
+        key_states = self._shape(key_states, -1, bsz)
+        value_states = self._shape(value_states, -1, bsz)
+    elif past_key_value is not None:
+        kv_states = self.kv_proj.kv_forward(hidden_states)
+        kv_states = torch.cat([past_key_value[0], kv_states], dim=1)
+        key_states, value_states = self.kv_proj.kv_up(kv_states)
+        key_states = self._shape(key_states, -1, bsz)
+        value_states = self._shape(value_states, -1, bsz)
+    else:
+        kv_states = self.kv_proj.kv_forward(hidden_states)
+        key_states, value_states = self.kv_proj.kv_up(kv_states)
+        key_states = self._shape(key_states, -1, bsz)
+        value_states = self._shape(value_states, -1, bsz)
+        
+    # NOTE: the code below has not been modified.
+    if (self.is_decoder) and not is_cross_attention:
+            # if cross_attention save Tuple(torch.Tensor, torch.Tensor) of all cross attention key/value_states.
+            # Further calls to cross_attention layer can then reuse all cross-attention
+            # key/value_states (first "if" case)
+            # if uni-directional self-attention (decoder) save Tuple(torch.Tensor, torch.Tensor) of
+            # all previous decoder key/value_states. Further calls to uni-directional self-attention
+            # can concat previous decoder key/value_states to current projected key/value_states (third "elif" case)
+            # if encoder bi-directional self-attention `past_key_value` is always `None`
+        past_key_value = (kv_states, )
+
+    proj_shape = (bsz * self.num_heads, -1, self.head_dim)
+    query_states = self._shape(query_states, tgt_len, bsz).view(*proj_shape)
+    key_states = key_states.reshape(*proj_shape)
+    value_states = value_states.reshape(*proj_shape)
+
+    src_len = key_states.size(1)
+    attn_weights = torch.bmm(query_states, key_states.transpose(1, 2))
+
+    if attn_weights.size() != (bsz * self.num_heads, tgt_len, src_len):
+        raise ValueError(
+            f"Attention weights should be of size {(bsz * self.num_heads, tgt_len, src_len)}, but is"
+            f" {attn_weights.size()}"
+        )
+
+    if attention_mask is not None:
+        if attention_mask.size() != (bsz, 1, tgt_len, src_len):
+            raise ValueError(
+                f"Attention mask should be of size {(bsz, 1, tgt_len, src_len)}, but is {attention_mask.size()}"
+                )
+        attn_weights = attn_weights.view(bsz, self.num_heads, tgt_len, src_len) + attention_mask
+        attn_weights = attn_weights.view(bsz * self.num_heads, tgt_len, src_len)
+
+    attn_weights = nn.functional.softmax(attn_weights, dim=-1)
+
+    if layer_head_mask is not None:
+        if layer_head_mask.size() != (self.num_heads,):
+            raise ValueError(
+                f"Head mask for a single layer should be of size {(self.num_heads,)}, but is"
+                f" {layer_head_mask.size()}"
+            )
+        attn_weights = layer_head_mask.view(1, -1, 1, 1) * attn_weights.view(bsz, self.num_heads, tgt_len, src_len)
+        attn_weights = attn_weights.view(bsz * self.num_heads, tgt_len, src_len)
+
+    if output_attentions:
+        attn_weights_reshaped = attn_weights.view(bsz, self.num_heads, tgt_len, src_len)
+        attn_weights = attn_weights_reshaped.view(bsz * self.num_heads, tgt_len, src_len)
+    else:
+        attn_weights_reshaped = None
+
+    attn_probs = nn.functional.dropout(attn_weights, p=self.dropout, training=self.training)
+
+    attn_output = torch.bmm(attn_probs, value_states)
+
+    if attn_output.size() != (bsz * self.num_heads, tgt_len, self.head_dim):
+        raise ValueError(
+            f"`attn_output` should be of size {(bsz * self.num_heads, tgt_len, self.head_dim)}, but is"
+            f" {attn_output.size()}"
+        )
+
+    attn_output = attn_output.view(bsz, self.num_heads, tgt_len, self.head_dim)
+    attn_output = attn_output.transpose(1, 2)
+
+    attn_output = attn_output.reshape(bsz, tgt_len, self.embed_dim)
+
+    attn_output = self.out_proj(attn_output)
+   
+    return attn_output, attn_weights_reshaped, past_key_value
+
+
+'''def mha2mla_mla_whisper(model):
+    for module in model.modules():
+        if isinstance(module, WhisperDecoderLayer):
+            module.self_attn.forward = types.MethodType(
+                custom_WhisperAttention_mla_forward,
+                module.self_attn
+            )'''
+def mha2mla_mla_whisper(_model=None):
+    WhisperAttention.forward = custom_WhisperAttention_nosplit_forward

preprocessor_config.json

@@ -0,0 +1,14 @@
+{
+  "chunk_length": 30,
+  "feature_extractor_type": "WhisperFeatureExtractor",
+  "feature_size": 80,
+  "hop_length": 160,
+  "n_fft": 400,
+  "n_samples": 480000,
+  "nb_max_frames": 3000,
+  "padding_side": "right",
+  "padding_value": 0.0,
+  "processor_class": "WhisperProcessor",
+  "return_attention_mask": false,
+  "sampling_rate": 16000
+}

special_tokens_map.json

@@ -0,0 +1,139 @@
+{
+  "additional_special_tokens": [
+    "<|endoftext|>",
+    "<|startoftranscript|>",
+    "<|en|>",
+    "<|zh|>",
+    "<|de|>",
+    "<|es|>",
+    "<|ru|>",
+    "<|ko|>",
+    "<|fr|>",
+    "<|ja|>",
+    "<|pt|>",
+    "<|tr|>",
+    "<|pl|>",
+    "<|ca|>",
+    "<|nl|>",
+    "<|ar|>",
+    "<|sv|>",
+    "<|it|>",
+    "<|id|>",
+    "<|hi|>",
+    "<|fi|>",
+    "<|vi|>",
+    "<|he|>",
+    "<|uk|>",
+    "<|el|>",
+    "<|ms|>",
+    "<|cs|>",
+    "<|ro|>",
+    "<|da|>",
+    "<|hu|>",
+    "<|ta|>",
+    "<|no|>",
+    "<|th|>",
+    "<|ur|>",
+    "<|hr|>",
+    "<|bg|>",
+    "<|lt|>",
+    "<|la|>",
+    "<|mi|>",
+    "<|ml|>",
+    "<|cy|>",
+    "<|sk|>",
+    "<|te|>",
+    "<|fa|>",
+    "<|lv|>",
+    "<|bn|>",
+    "<|sr|>",
+    "<|az|>",
+    "<|sl|>",
+    "<|kn|>",
+    "<|et|>",
+    "<|mk|>",
+    "<|br|>",
+    "<|eu|>",
+    "<|is|>",
+    "<|hy|>",
+    "<|ne|>",
+    "<|mn|>",
+    "<|bs|>",
+    "<|kk|>",
+    "<|sq|>",
+    "<|sw|>",
+    "<|gl|>",
+    "<|mr|>",
+    "<|pa|>",
+    "<|si|>",
+    "<|km|>",
+    "<|sn|>",
+    "<|yo|>",
+    "<|so|>",
+    "<|af|>",
+    "<|oc|>",
+    "<|ka|>",
+    "<|be|>",
+    "<|tg|>",
+    "<|sd|>",
+    "<|gu|>",
+    "<|am|>",
+    "<|yi|>",
+    "<|lo|>",
+    "<|uz|>",
+    "<|fo|>",
+    "<|ht|>",
+    "<|ps|>",
+    "<|tk|>",
+    "<|nn|>",
+    "<|mt|>",
+    "<|sa|>",
+    "<|lb|>",
+    "<|my|>",
+    "<|bo|>",
+    "<|tl|>",
+    "<|mg|>",
+    "<|as|>",
+    "<|tt|>",
+    "<|haw|>",
+    "<|ln|>",
+    "<|ha|>",
+    "<|ba|>",
+    "<|jw|>",
+    "<|su|>",
+    "<|translate|>",
+    "<|transcribe|>",
+    "<|startoflm|>",
+    "<|startofprev|>",
+    "<|nocaptions|>",
+    "<|notimestamps|>"
+  ],
+  "bos_token": {
+    "content": "<|endoftext|>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "eos_token": {
+    "content": "<|endoftext|>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "pad_token": {
+    "content": "<|endoftext|>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "unk_token": {
+    "content": "<|endoftext|>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  }
+}