Fix config.json (MTP layer was claimed as NVFP4 by layer-agnostic targets); correct MTP guidance with measured 56.5% acceptance; add unquantized MoE oracle patch
Browse files- README.md +55 -0
- config.json +1536 -0
- vllm_patch/README.md +21 -0
- vllm_patch/unquantized_moe_oracle.py +501 -0
README.md
CHANGED
|
@@ -21,12 +21,67 @@ produced with [**qstream**](https://github.com/olka/qstream). The routed MoE exp
|
|
| 21 |
|
| 22 |
**The original model card follows in full [below](#original-model-card).**
|
| 23 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 24 |
| | |
|
| 25 |
|---|---|
|
| 26 |
| **Size** | **81.4 GB** (down from 255.0 GB BF16 source, ~32%) |
|
| 27 |
| **Format** | compressed-tensors `nvfp4-pack-quantized` (E2M1 4-bit + FP8-E4M3 group-16 scales + per-tensor global scale) |
|
| 28 |
| **Base** | Ling-3.0-flash — 124B total / 5.1B active hybrid-linear MoE; 42 layers stacked 5:1 as 35 Kimi-Delta-Attention (KDA) + 7 gated-MLA; 512 routed experts top-8 + 1 shared; 2 dense layers; 1 MTP layer; 256K context |
|
| 29 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 30 |
|
| 31 |
## What is quantized to what
|
| 32 |
|
|
|
|
| 21 |
|
| 22 |
**The original model card follows in full [below](#original-model-card).**
|
| 23 |
|
| 24 |
+
> ### ⚠️ Speculative decoding (MTP) needs specific flags **and** a vLLM patch
|
| 25 |
+
>
|
| 26 |
+
> It works well — **56.5% acceptance, 2.70 tokens per decode step** (94.3% at draft
|
| 27 |
+
> position 0) — but NVFP4 needs more setup than the
|
| 28 |
+
> [MXFP4 build](https://huggingface.co/olka-fi/Ling-3.0-flash-MXFP4) does:
|
| 29 |
+
>
|
| 30 |
+
> ```
|
| 31 |
+
> --kernel-config '{"moe_backend":"marlin"}' --max-num-seqs 256
|
| 32 |
+
> ```
|
| 33 |
+
>
|
| 34 |
+
> plus the `unquantized.py` patch in [`vllm_patch/`](./vllm_patch/). Why each is needed:
|
| 35 |
+
>
|
| 36 |
+
> * **`marlin`** is the only MoE backend supporting weight-only NVFP4A16. `triton` is
|
| 37 |
+
> rejected outright; `cutlass` and `flashinfer_cutlass` reject the scheme
|
| 38 |
+
> (`QuantKey(u8,scale(f8e4m3fn,...))`).
|
| 39 |
+
> * **The patch** exists because `moe_backend` is a single *global* setting, while this
|
| 40 |
+
> checkpoint has two MoE kinds: NVFP4 routed experts (need marlin) and the **BF16 MTP
|
| 41 |
+
> layer** (marlin has no unquantized kernel → `moe_backend='marlin' is not supported
|
| 42 |
+
> for unquantized MoE`). The patch routes unquantized MoE to Triton, mirroring the
|
| 43 |
+
> existing `humming` precedent. Triton also avoids FlashInfer TRT-LLM's grouped-routing
|
| 44 |
+
> kernel, which is warp-limited to 32 experts per group — Ling-3.0 has **64** (512
|
| 45 |
+
> experts / 8 groups) and either crashes there or silently mis-routes.
|
| 46 |
+
> * **`--max-num-seqs 256`** — the BF16 MTP head takes ~6 GB, shrinking the KDA state
|
| 47 |
+
> cache below the default 1024 sequences, so CUDA graph capture aborts with
|
| 48 |
+
> `max_num_seqs (1024) exceeds available Mamba cache blocks (684)`.
|
| 49 |
+
>
|
| 50 |
+
> Acceptance by draft position (conditional): **94.3%** / 57.2% / 39.7%. The falloff is
|
| 51 |
+
> expected — Ling-3.0 has a single MTP layer re-run per speculative token — so
|
| 52 |
+
> `num_speculative_tokens: 2` may beat 3 on net throughput. MTP also roughly doubles cold
|
| 53 |
+
> start: the draft loader re-reads the whole checkpoint to extract one layer.
|
| 54 |
+
>
|
| 55 |
+
> With all of the above, graph capture succeeds and `--enforce-eager` is *not* needed.
|
| 56 |
+
> **If you served this checkpoint before 2026-08-05, re-pull `config.json`** — the earlier
|
| 57 |
+
> one let the layer-agnostic targets regex claim the MTP layer's BF16 experts as NVFP4,
|
| 58 |
+
> giving a silently broken drafter (0% acceptance, no error).
|
| 59 |
+
|
| 60 |
| | |
|
| 61 |
|---|---|
|
| 62 |
| **Size** | **81.4 GB** (down from 255.0 GB BF16 source, ~32%) |
|
| 63 |
| **Format** | compressed-tensors `nvfp4-pack-quantized` (E2M1 4-bit + FP8-E4M3 group-16 scales + per-tensor global scale) |
|
| 64 |
| **Base** | Ling-3.0-flash — 124B total / 5.1B active hybrid-linear MoE; 42 layers stacked 5:1 as 35 Kimi-Delta-Attention (KDA) + 7 gated-MLA; 512 routed experts top-8 + 1 shared; 2 dense layers; 1 MTP layer; 256K context |
|
| 65 |
|
| 66 |
+
## Which build should you use?
|
| 67 |
+
|
| 68 |
+
**Most people should take the [MXFP4 build](https://huggingface.co/olka-fi/Ling-3.0-flash-MXFP4) instead.**
|
| 69 |
+
NVFP4 reconstructs the weights substantially more faithfully, and that advantage does not
|
| 70 |
+
show up downstream:
|
| 71 |
+
|
| 72 |
+
| | [MXFP4](https://huggingface.co/olka-fi/Ling-3.0-flash-MXFP4) | NVFP4 (this repo) |
|
| 73 |
+
|---|---|---|
|
| 74 |
+
| Size | **77.6 GB** | 81.4 GB |
|
| 75 |
+
| Median weight rel. error | 0.111618 | **0.086498** (−22.5%) |
|
| 76 |
+
| GSM8K 5-shot, full 1319, flexible-extract | **84.38%** ±1.00 | 83.17% ±1.03 |
|
| 77 |
+
| GSM8K 5-shot, full 1319, strict-match | **78.85%** ±1.12 | 78.39% ±1.13 |
|
| 78 |
+
|
| 79 |
+
Both GSM8K gaps are under 1.2σ — the two builds are statistically indistinguishable on
|
| 80 |
+
this benchmark despite NVFP4's much lower weight error. At this magnitude, 4-bit
|
| 81 |
+
reconstruction error is already below what GSM8K can resolve. NVFP4 is published for
|
| 82 |
+
comparison and for anyone wanting the more faithful weights (e.g. for longer-generation
|
| 83 |
+
or harder tasks where the difference may still surface); MXFP4 is the smaller, better-
|
| 84 |
+
tested release.
|
| 85 |
|
| 86 |
## What is quantized to what
|
| 87 |
|
config.json
CHANGED
|
@@ -207,6 +207,1542 @@
|
|
| 207 |
"re:.*model\\.layers\\.\\d+\\.attention\\.kv_a_proj_with_mqa$",
|
| 208 |
"re:.*model\\.layers\\.\\d+\\.attention\\.kv_b_proj$",
|
| 209 |
"re:.*model\\.layers\\.\\d+\\.eh_proj$",
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 210 |
"re:.*model\\.word_embeddings$"
|
| 211 |
]
|
| 212 |
}
|
|
|
|
| 207 |
"re:.*model\\.layers\\.\\d+\\.attention\\.kv_a_proj_with_mqa$",
|
| 208 |
"re:.*model\\.layers\\.\\d+\\.attention\\.kv_b_proj$",
|
| 209 |
"re:.*model\\.layers\\.\\d+\\.eh_proj$",
|
| 210 |
+
"model.layers.42.mlp.experts.0.down_proj",
|
| 211 |
+
"model.layers.42.mlp.experts.1.down_proj",
|
| 212 |
+
"model.layers.42.mlp.experts.10.down_proj",
|
| 213 |
+
"model.layers.42.mlp.experts.100.down_proj",
|
| 214 |
+
"model.layers.42.mlp.experts.101.down_proj",
|
| 215 |
+
"model.layers.42.mlp.experts.102.down_proj",
|
| 216 |
+
"model.layers.42.mlp.experts.103.down_proj",
|
| 217 |
+
"model.layers.42.mlp.experts.104.down_proj",
|
| 218 |
+
"model.layers.42.mlp.experts.105.down_proj",
|
| 219 |
+
"model.layers.42.mlp.experts.106.down_proj",
|
| 220 |
+
"model.layers.42.mlp.experts.107.down_proj",
|
| 221 |
+
"model.layers.42.mlp.experts.108.down_proj",
|
| 222 |
+
"model.layers.42.mlp.experts.109.down_proj",
|
| 223 |
+
"model.layers.42.mlp.experts.11.down_proj",
|
| 224 |
+
"model.layers.42.mlp.experts.110.down_proj",
|
| 225 |
+
"model.layers.42.mlp.experts.111.down_proj",
|
| 226 |
+
"model.layers.42.mlp.experts.112.down_proj",
|
| 227 |
+
"model.layers.42.mlp.experts.113.down_proj",
|
| 228 |
+
"model.layers.42.mlp.experts.114.down_proj",
|
| 229 |
+
"model.layers.42.mlp.experts.115.down_proj",
|
| 230 |
+
"model.layers.42.mlp.experts.116.down_proj",
|
| 231 |
+
"model.layers.42.mlp.experts.117.down_proj",
|
| 232 |
+
"model.layers.42.mlp.experts.118.down_proj",
|
| 233 |
+
"model.layers.42.mlp.experts.119.down_proj",
|
| 234 |
+
"model.layers.42.mlp.experts.12.down_proj",
|
| 235 |
+
"model.layers.42.mlp.experts.120.down_proj",
|
| 236 |
+
"model.layers.42.mlp.experts.121.down_proj",
|
| 237 |
+
"model.layers.42.mlp.experts.122.down_proj",
|
| 238 |
+
"model.layers.42.mlp.experts.123.down_proj",
|
| 239 |
+
"model.layers.42.mlp.experts.124.down_proj",
|
| 240 |
+
"model.layers.42.mlp.experts.125.down_proj",
|
| 241 |
+
"model.layers.42.mlp.experts.126.down_proj",
|
| 242 |
+
"model.layers.42.mlp.experts.127.down_proj",
|
| 243 |
+
"model.layers.42.mlp.experts.128.down_proj",
|
| 244 |
+
"model.layers.42.mlp.experts.129.down_proj",
|
| 245 |
+
"model.layers.42.mlp.experts.13.down_proj",
|
| 246 |
+
"model.layers.42.mlp.experts.130.down_proj",
|
| 247 |
+
"model.layers.42.mlp.experts.131.down_proj",
|
| 248 |
+
"model.layers.42.mlp.experts.132.down_proj",
|
| 249 |
+
"model.layers.42.mlp.experts.133.down_proj",
|
| 250 |
+
"model.layers.42.mlp.experts.134.down_proj",
|
| 251 |
+
"model.layers.42.mlp.experts.135.down_proj",
|
| 252 |
+
"model.layers.42.mlp.experts.136.down_proj",
|
| 253 |
+
"model.layers.42.mlp.experts.137.down_proj",
|
| 254 |
+
"model.layers.42.mlp.experts.138.down_proj",
|
| 255 |
+
"model.layers.42.mlp.experts.139.down_proj",
|
| 256 |
+
"model.layers.42.mlp.experts.14.down_proj",
|
| 257 |
+
"model.layers.42.mlp.experts.140.down_proj",
|
| 258 |
+
"model.layers.42.mlp.experts.141.down_proj",
|
| 259 |
+
"model.layers.42.mlp.experts.142.down_proj",
|
| 260 |
+
"model.layers.42.mlp.experts.143.down_proj",
|
| 261 |
+
"model.layers.42.mlp.experts.144.down_proj",
|
| 262 |
+
"model.layers.42.mlp.experts.145.down_proj",
|
| 263 |
+
"model.layers.42.mlp.experts.146.down_proj",
|
| 264 |
+
"model.layers.42.mlp.experts.147.down_proj",
|
| 265 |
+
"model.layers.42.mlp.experts.148.down_proj",
|
| 266 |
+
"model.layers.42.mlp.experts.149.down_proj",
|
| 267 |
+
"model.layers.42.mlp.experts.15.down_proj",
|
| 268 |
+
"model.layers.42.mlp.experts.150.down_proj",
|
| 269 |
+
"model.layers.42.mlp.experts.151.down_proj",
|
| 270 |
+
"model.layers.42.mlp.experts.152.down_proj",
|
| 271 |
+
"model.layers.42.mlp.experts.153.down_proj",
|
| 272 |
+
"model.layers.42.mlp.experts.154.down_proj",
|
| 273 |
+
"model.layers.42.mlp.experts.155.down_proj",
|
| 274 |
+
"model.layers.42.mlp.experts.156.down_proj",
|
| 275 |
+
"model.layers.42.mlp.experts.157.down_proj",
|
| 276 |
+
"model.layers.42.mlp.experts.158.down_proj",
|
| 277 |
+
"model.layers.42.mlp.experts.159.down_proj",
|
| 278 |
+
"model.layers.42.mlp.experts.16.down_proj",
|
| 279 |
+
"model.layers.42.mlp.experts.160.down_proj",
|
| 280 |
+
"model.layers.42.mlp.experts.161.down_proj",
|
| 281 |
+
"model.layers.42.mlp.experts.162.down_proj",
|
| 282 |
+
"model.layers.42.mlp.experts.163.down_proj",
|
| 283 |
+
"model.layers.42.mlp.experts.164.down_proj",
|
| 284 |
+
"model.layers.42.mlp.experts.165.down_proj",
|
| 285 |
+
"model.layers.42.mlp.experts.166.down_proj",
|
| 286 |
+
"model.layers.42.mlp.experts.167.down_proj",
|
| 287 |
+
"model.layers.42.mlp.experts.168.down_proj",
|
| 288 |
+
"model.layers.42.mlp.experts.169.down_proj",
|
| 289 |
+
"model.layers.42.mlp.experts.17.down_proj",
|
| 290 |
+
"model.layers.42.mlp.experts.170.down_proj",
|
| 291 |
+
"model.layers.42.mlp.experts.171.down_proj",
|
| 292 |
+
"model.layers.42.mlp.experts.172.down_proj",
|
| 293 |
+
"model.layers.42.mlp.experts.173.down_proj",
|
| 294 |
+
"model.layers.42.mlp.experts.174.down_proj",
|
| 295 |
+
"model.layers.42.mlp.experts.175.down_proj",
|
| 296 |
+
"model.layers.42.mlp.experts.176.down_proj",
|
| 297 |
+
"model.layers.42.mlp.experts.177.down_proj",
|
| 298 |
+
"model.layers.42.mlp.experts.178.down_proj",
|
| 299 |
+
"model.layers.42.mlp.experts.179.down_proj",
|
| 300 |
+
"model.layers.42.mlp.experts.18.down_proj",
|
| 301 |
+
"model.layers.42.mlp.experts.180.down_proj",
|
| 302 |
+
"model.layers.42.mlp.experts.181.down_proj",
|
| 303 |
+
"model.layers.42.mlp.experts.182.down_proj",
|
| 304 |
+
"model.layers.42.mlp.experts.183.down_proj",
|
| 305 |
+
"model.layers.42.mlp.experts.184.down_proj",
|
| 306 |
+
"model.layers.42.mlp.experts.185.down_proj",
|
| 307 |
+
"model.layers.42.mlp.experts.186.down_proj",
|
| 308 |
+
"model.layers.42.mlp.experts.187.down_proj",
|
| 309 |
+
"model.layers.42.mlp.experts.188.down_proj",
|
| 310 |
+
"model.layers.42.mlp.experts.189.down_proj",
|
| 311 |
+
"model.layers.42.mlp.experts.19.down_proj",
|
| 312 |
+
"model.layers.42.mlp.experts.190.down_proj",
|
| 313 |
+
"model.layers.42.mlp.experts.191.down_proj",
|
| 314 |
+
"model.layers.42.mlp.experts.192.down_proj",
|
| 315 |
+
"model.layers.42.mlp.experts.193.down_proj",
|
| 316 |
+
"model.layers.42.mlp.experts.194.down_proj",
|
| 317 |
+
"model.layers.42.mlp.experts.195.down_proj",
|
| 318 |
+
"model.layers.42.mlp.experts.196.down_proj",
|
| 319 |
+
"model.layers.42.mlp.experts.197.down_proj",
|
| 320 |
+
"model.layers.42.mlp.experts.198.down_proj",
|
| 321 |
+
"model.layers.42.mlp.experts.199.down_proj",
|
| 322 |
+
"model.layers.42.mlp.experts.2.down_proj",
|
| 323 |
+
"model.layers.42.mlp.experts.20.down_proj",
|
| 324 |
+
"model.layers.42.mlp.experts.200.down_proj",
|
| 325 |
+
"model.layers.42.mlp.experts.201.down_proj",
|
| 326 |
+
"model.layers.42.mlp.experts.202.down_proj",
|
| 327 |
+
"model.layers.42.mlp.experts.203.down_proj",
|
| 328 |
+
"model.layers.42.mlp.experts.204.down_proj",
|
| 329 |
+
"model.layers.42.mlp.experts.205.down_proj",
|
| 330 |
+
"model.layers.42.mlp.experts.206.down_proj",
|
| 331 |
+
"model.layers.42.mlp.experts.207.down_proj",
|
| 332 |
+
"model.layers.42.mlp.experts.208.down_proj",
|
| 333 |
+
"model.layers.42.mlp.experts.209.down_proj",
|
| 334 |
+
"model.layers.42.mlp.experts.21.down_proj",
|
| 335 |
+
"model.layers.42.mlp.experts.210.down_proj",
|
| 336 |
+
"model.layers.42.mlp.experts.211.down_proj",
|
| 337 |
+
"model.layers.42.mlp.experts.212.down_proj",
|
| 338 |
+
"model.layers.42.mlp.experts.213.down_proj",
|
| 339 |
+
"model.layers.42.mlp.experts.214.down_proj",
|
| 340 |
+
"model.layers.42.mlp.experts.215.down_proj",
|
| 341 |
+
"model.layers.42.mlp.experts.216.down_proj",
|
| 342 |
+
"model.layers.42.mlp.experts.217.down_proj",
|
| 343 |
+
"model.layers.42.mlp.experts.218.down_proj",
|
| 344 |
+
"model.layers.42.mlp.experts.219.down_proj",
|
| 345 |
+
"model.layers.42.mlp.experts.22.down_proj",
|
| 346 |
+
"model.layers.42.mlp.experts.220.down_proj",
|
| 347 |
+
"model.layers.42.mlp.experts.221.down_proj",
|
| 348 |
+
"model.layers.42.mlp.experts.222.down_proj",
|
| 349 |
+
"model.layers.42.mlp.experts.223.down_proj",
|
| 350 |
+
"model.layers.42.mlp.experts.224.down_proj",
|
| 351 |
+
"model.layers.42.mlp.experts.225.down_proj",
|
| 352 |
+
"model.layers.42.mlp.experts.226.down_proj",
|
| 353 |
+
"model.layers.42.mlp.experts.227.down_proj",
|
| 354 |
+
"model.layers.42.mlp.experts.228.down_proj",
|
| 355 |
+
"model.layers.42.mlp.experts.229.down_proj",
|
| 356 |
+
"model.layers.42.mlp.experts.23.down_proj",
|
| 357 |
+
"model.layers.42.mlp.experts.230.down_proj",
|
| 358 |
+
"model.layers.42.mlp.experts.231.down_proj",
|
| 359 |
+
"model.layers.42.mlp.experts.232.down_proj",
|
| 360 |
+
"model.layers.42.mlp.experts.233.down_proj",
|
| 361 |
+
"model.layers.42.mlp.experts.234.down_proj",
|
| 362 |
+
"model.layers.42.mlp.experts.235.down_proj",
|
| 363 |
+
"model.layers.42.mlp.experts.236.down_proj",
|
| 364 |
+
"model.layers.42.mlp.experts.237.down_proj",
|
| 365 |
+
"model.layers.42.mlp.experts.238.down_proj",
|
| 366 |
+
"model.layers.42.mlp.experts.239.down_proj",
|
| 367 |
+
"model.layers.42.mlp.experts.24.down_proj",
|
| 368 |
+
"model.layers.42.mlp.experts.240.down_proj",
|
| 369 |
+
"model.layers.42.mlp.experts.241.down_proj",
|
| 370 |
+
"model.layers.42.mlp.experts.242.down_proj",
|
| 371 |
+
"model.layers.42.mlp.experts.243.down_proj",
|
| 372 |
+
"model.layers.42.mlp.experts.244.down_proj",
|
| 373 |
+
"model.layers.42.mlp.experts.245.down_proj",
|
| 374 |
+
"model.layers.42.mlp.experts.246.down_proj",
|
| 375 |
+
"model.layers.42.mlp.experts.247.down_proj",
|
| 376 |
+
"model.layers.42.mlp.experts.248.down_proj",
|
| 377 |
+
"model.layers.42.mlp.experts.249.down_proj",
|
| 378 |
+
"model.layers.42.mlp.experts.25.down_proj",
|
| 379 |
+
"model.layers.42.mlp.experts.250.down_proj",
|
| 380 |
+
"model.layers.42.mlp.experts.251.down_proj",
|
| 381 |
+
"model.layers.42.mlp.experts.252.down_proj",
|
| 382 |
+
"model.layers.42.mlp.experts.253.down_proj",
|
| 383 |
+
"model.layers.42.mlp.experts.254.down_proj",
|
| 384 |
+
"model.layers.42.mlp.experts.255.down_proj",
|
| 385 |
+
"model.layers.42.mlp.experts.256.down_proj",
|
| 386 |
+
"model.layers.42.mlp.experts.257.down_proj",
|
| 387 |
+
"model.layers.42.mlp.experts.258.down_proj",
|
| 388 |
+
"model.layers.42.mlp.experts.259.down_proj",
|
| 389 |
+
"model.layers.42.mlp.experts.26.down_proj",
|
| 390 |
+
"model.layers.42.mlp.experts.260.down_proj",
|
| 391 |
+
"model.layers.42.mlp.experts.261.down_proj",
|
| 392 |
+
"model.layers.42.mlp.experts.262.down_proj",
|
| 393 |
+
"model.layers.42.mlp.experts.263.down_proj",
|
| 394 |
+
"model.layers.42.mlp.experts.264.down_proj",
|
| 395 |
+
"model.layers.42.mlp.experts.265.down_proj",
|
| 396 |
+
"model.layers.42.mlp.experts.266.down_proj",
|
| 397 |
+
"model.layers.42.mlp.experts.267.down_proj",
|
| 398 |
+
"model.layers.42.mlp.experts.268.down_proj",
|
| 399 |
+
"model.layers.42.mlp.experts.269.down_proj",
|
| 400 |
+
"model.layers.42.mlp.experts.27.down_proj",
|
| 401 |
+
"model.layers.42.mlp.experts.270.down_proj",
|
| 402 |
+
"model.layers.42.mlp.experts.271.down_proj",
|
| 403 |
+
"model.layers.42.mlp.experts.272.down_proj",
|
| 404 |
+
"model.layers.42.mlp.experts.273.down_proj",
|
| 405 |
+
"model.layers.42.mlp.experts.274.down_proj",
|
| 406 |
+
"model.layers.42.mlp.experts.275.down_proj",
|
| 407 |
+
"model.layers.42.mlp.experts.276.down_proj",
|
| 408 |
+
"model.layers.42.mlp.experts.277.down_proj",
|
| 409 |
+
"model.layers.42.mlp.experts.278.down_proj",
|
| 410 |
+
"model.layers.42.mlp.experts.279.down_proj",
|
| 411 |
+
"model.layers.42.mlp.experts.28.down_proj",
|
| 412 |
+
"model.layers.42.mlp.experts.280.down_proj",
|
| 413 |
+
"model.layers.42.mlp.experts.281.down_proj",
|
| 414 |
+
"model.layers.42.mlp.experts.282.down_proj",
|
| 415 |
+
"model.layers.42.mlp.experts.283.down_proj",
|
| 416 |
+
"model.layers.42.mlp.experts.284.down_proj",
|
| 417 |
+
"model.layers.42.mlp.experts.285.down_proj",
|
| 418 |
+
"model.layers.42.mlp.experts.286.down_proj",
|
| 419 |
+
"model.layers.42.mlp.experts.287.down_proj",
|
| 420 |
+
"model.layers.42.mlp.experts.288.down_proj",
|
| 421 |
+
"model.layers.42.mlp.experts.289.down_proj",
|
| 422 |
+
"model.layers.42.mlp.experts.29.down_proj",
|
| 423 |
+
"model.layers.42.mlp.experts.290.down_proj",
|
| 424 |
+
"model.layers.42.mlp.experts.291.down_proj",
|
| 425 |
+
"model.layers.42.mlp.experts.292.down_proj",
|
| 426 |
+
"model.layers.42.mlp.experts.293.down_proj",
|
| 427 |
+
"model.layers.42.mlp.experts.294.down_proj",
|
| 428 |
+
"model.layers.42.mlp.experts.295.down_proj",
|
| 429 |
+
"model.layers.42.mlp.experts.296.down_proj",
|
| 430 |
+
"model.layers.42.mlp.experts.297.down_proj",
|
| 431 |
+
"model.layers.42.mlp.experts.298.down_proj",
|
| 432 |
+
"model.layers.42.mlp.experts.299.down_proj",
|
| 433 |
+
"model.layers.42.mlp.experts.3.down_proj",
|
| 434 |
+
"model.layers.42.mlp.experts.30.down_proj",
|
| 435 |
+
"model.layers.42.mlp.experts.300.down_proj",
|
| 436 |
+
"model.layers.42.mlp.experts.301.down_proj",
|
| 437 |
+
"model.layers.42.mlp.experts.302.down_proj",
|
| 438 |
+
"model.layers.42.mlp.experts.303.down_proj",
|
| 439 |
+
"model.layers.42.mlp.experts.304.down_proj",
|
| 440 |
+
"model.layers.42.mlp.experts.305.down_proj",
|
| 441 |
+
"model.layers.42.mlp.experts.306.down_proj",
|
| 442 |
+
"model.layers.42.mlp.experts.307.down_proj",
|
| 443 |
+
"model.layers.42.mlp.experts.308.down_proj",
|
| 444 |
+
"model.layers.42.mlp.experts.309.down_proj",
|
| 445 |
+
"model.layers.42.mlp.experts.31.down_proj",
|
| 446 |
+
"model.layers.42.mlp.experts.310.down_proj",
|
| 447 |
+
"model.layers.42.mlp.experts.311.down_proj",
|
| 448 |
+
"model.layers.42.mlp.experts.312.down_proj",
|
| 449 |
+
"model.layers.42.mlp.experts.313.down_proj",
|
| 450 |
+
"model.layers.42.mlp.experts.314.down_proj",
|
| 451 |
+
"model.layers.42.mlp.experts.315.down_proj",
|
| 452 |
+
"model.layers.42.mlp.experts.316.down_proj",
|
| 453 |
+
"model.layers.42.mlp.experts.317.down_proj",
|
| 454 |
+
"model.layers.42.mlp.experts.318.down_proj",
|
| 455 |
+
"model.layers.42.mlp.experts.319.down_proj",
|
| 456 |
+
"model.layers.42.mlp.experts.32.down_proj",
|
| 457 |
+
"model.layers.42.mlp.experts.320.down_proj",
|
| 458 |
+
"model.layers.42.mlp.experts.321.down_proj",
|
| 459 |
+
"model.layers.42.mlp.experts.322.down_proj",
|
| 460 |
+
"model.layers.42.mlp.experts.323.down_proj",
|
| 461 |
+
"model.layers.42.mlp.experts.324.down_proj",
|
| 462 |
+
"model.layers.42.mlp.experts.325.down_proj",
|
| 463 |
+
"model.layers.42.mlp.experts.326.down_proj",
|
| 464 |
+
"model.layers.42.mlp.experts.327.down_proj",
|
| 465 |
+
"model.layers.42.mlp.experts.328.down_proj",
|
| 466 |
+
"model.layers.42.mlp.experts.329.down_proj",
|
| 467 |
+
"model.layers.42.mlp.experts.33.down_proj",
|
| 468 |
+
"model.layers.42.mlp.experts.330.down_proj",
|
| 469 |
+
"model.layers.42.mlp.experts.331.down_proj",
|
| 470 |
+
"model.layers.42.mlp.experts.332.down_proj",
|
| 471 |
+
"model.layers.42.mlp.experts.333.down_proj",
|
| 472 |
+
"model.layers.42.mlp.experts.334.down_proj",
|
| 473 |
+
"model.layers.42.mlp.experts.335.down_proj",
|
| 474 |
+
"model.layers.42.mlp.experts.336.down_proj",
|
| 475 |
+
"model.layers.42.mlp.experts.337.down_proj",
|
| 476 |
+
"model.layers.42.mlp.experts.338.down_proj",
|
| 477 |
+
"model.layers.42.mlp.experts.339.down_proj",
|
| 478 |
+
"model.layers.42.mlp.experts.34.down_proj",
|
| 479 |
+
"model.layers.42.mlp.experts.340.down_proj",
|
| 480 |
+
"model.layers.42.mlp.experts.341.down_proj",
|
| 481 |
+
"model.layers.42.mlp.experts.342.down_proj",
|
| 482 |
+
"model.layers.42.mlp.experts.343.down_proj",
|
| 483 |
+
"model.layers.42.mlp.experts.344.down_proj",
|
| 484 |
+
"model.layers.42.mlp.experts.345.down_proj",
|
| 485 |
+
"model.layers.42.mlp.experts.346.down_proj",
|
| 486 |
+
"model.layers.42.mlp.experts.347.down_proj",
|
| 487 |
+
"model.layers.42.mlp.experts.348.down_proj",
|
| 488 |
+
"model.layers.42.mlp.experts.349.down_proj",
|
| 489 |
+
"model.layers.42.mlp.experts.35.down_proj",
|
| 490 |
+
"model.layers.42.mlp.experts.350.down_proj",
|
| 491 |
+
"model.layers.42.mlp.experts.351.down_proj",
|
| 492 |
+
"model.layers.42.mlp.experts.352.down_proj",
|
| 493 |
+
"model.layers.42.mlp.experts.353.down_proj",
|
| 494 |
+
"model.layers.42.mlp.experts.354.down_proj",
|
| 495 |
+
"model.layers.42.mlp.experts.355.down_proj",
|
| 496 |
+
"model.layers.42.mlp.experts.356.down_proj",
|
| 497 |
+
"model.layers.42.mlp.experts.357.down_proj",
|
| 498 |
+
"model.layers.42.mlp.experts.358.down_proj",
|
| 499 |
+
"model.layers.42.mlp.experts.359.down_proj",
|
| 500 |
+
"model.layers.42.mlp.experts.36.down_proj",
|
| 501 |
+
"model.layers.42.mlp.experts.360.down_proj",
|
| 502 |
+
"model.layers.42.mlp.experts.361.down_proj",
|
| 503 |
+
"model.layers.42.mlp.experts.362.down_proj",
|
| 504 |
+
"model.layers.42.mlp.experts.363.down_proj",
|
| 505 |
+
"model.layers.42.mlp.experts.364.down_proj",
|
| 506 |
+
"model.layers.42.mlp.experts.365.down_proj",
|
| 507 |
+
"model.layers.42.mlp.experts.366.down_proj",
|
| 508 |
+
"model.layers.42.mlp.experts.367.down_proj",
|
| 509 |
+
"model.layers.42.mlp.experts.368.down_proj",
|
| 510 |
+
"model.layers.42.mlp.experts.369.down_proj",
|
| 511 |
+
"model.layers.42.mlp.experts.37.down_proj",
|
| 512 |
+
"model.layers.42.mlp.experts.370.down_proj",
|
| 513 |
+
"model.layers.42.mlp.experts.371.down_proj",
|
| 514 |
+
"model.layers.42.mlp.experts.372.down_proj",
|
| 515 |
+
"model.layers.42.mlp.experts.373.down_proj",
|
| 516 |
+
"model.layers.42.mlp.experts.374.down_proj",
|
| 517 |
+
"model.layers.42.mlp.experts.375.down_proj",
|
| 518 |
+
"model.layers.42.mlp.experts.376.down_proj",
|
| 519 |
+
"model.layers.42.mlp.experts.377.down_proj",
|
| 520 |
+
"model.layers.42.mlp.experts.378.down_proj",
|
| 521 |
+
"model.layers.42.mlp.experts.379.down_proj",
|
| 522 |
+
"model.layers.42.mlp.experts.38.down_proj",
|
| 523 |
+
"model.layers.42.mlp.experts.380.down_proj",
|
| 524 |
+
"model.layers.42.mlp.experts.381.down_proj",
|
| 525 |
+
"model.layers.42.mlp.experts.382.down_proj",
|
| 526 |
+
"model.layers.42.mlp.experts.383.down_proj",
|
| 527 |
+
"model.layers.42.mlp.experts.384.down_proj",
|
| 528 |
+
"model.layers.42.mlp.experts.385.down_proj",
|
| 529 |
+
"model.layers.42.mlp.experts.386.down_proj",
|
| 530 |
+
"model.layers.42.mlp.experts.387.down_proj",
|
| 531 |
+
"model.layers.42.mlp.experts.388.down_proj",
|
| 532 |
+
"model.layers.42.mlp.experts.389.down_proj",
|
| 533 |
+
"model.layers.42.mlp.experts.39.down_proj",
|
| 534 |
+
"model.layers.42.mlp.experts.390.down_proj",
|
| 535 |
+
"model.layers.42.mlp.experts.391.down_proj",
|
| 536 |
+
"model.layers.42.mlp.experts.392.down_proj",
|
| 537 |
+
"model.layers.42.mlp.experts.393.down_proj",
|
| 538 |
+
"model.layers.42.mlp.experts.394.down_proj",
|
| 539 |
+
"model.layers.42.mlp.experts.395.down_proj",
|
| 540 |
+
"model.layers.42.mlp.experts.396.down_proj",
|
| 541 |
+
"model.layers.42.mlp.experts.397.down_proj",
|
| 542 |
+
"model.layers.42.mlp.experts.398.down_proj",
|
| 543 |
+
"model.layers.42.mlp.experts.399.down_proj",
|
| 544 |
+
"model.layers.42.mlp.experts.4.down_proj",
|
| 545 |
+
"model.layers.42.mlp.experts.40.down_proj",
|
| 546 |
+
"model.layers.42.mlp.experts.400.down_proj",
|
| 547 |
+
"model.layers.42.mlp.experts.401.down_proj",
|
| 548 |
+
"model.layers.42.mlp.experts.402.down_proj",
|
| 549 |
+
"model.layers.42.mlp.experts.403.down_proj",
|
| 550 |
+
"model.layers.42.mlp.experts.404.down_proj",
|
| 551 |
+
"model.layers.42.mlp.experts.405.down_proj",
|
| 552 |
+
"model.layers.42.mlp.experts.406.down_proj",
|
| 553 |
+
"model.layers.42.mlp.experts.407.down_proj",
|
| 554 |
+
"model.layers.42.mlp.experts.408.down_proj",
|
| 555 |
+
"model.layers.42.mlp.experts.409.down_proj",
|
| 556 |
+
"model.layers.42.mlp.experts.41.down_proj",
|
| 557 |
+
"model.layers.42.mlp.experts.410.down_proj",
|
| 558 |
+
"model.layers.42.mlp.experts.411.down_proj",
|
| 559 |
+
"model.layers.42.mlp.experts.412.down_proj",
|
| 560 |
+
"model.layers.42.mlp.experts.413.down_proj",
|
| 561 |
+
"model.layers.42.mlp.experts.414.down_proj",
|
| 562 |
+
"model.layers.42.mlp.experts.415.down_proj",
|
| 563 |
+
"model.layers.42.mlp.experts.416.down_proj",
|
| 564 |
+
"model.layers.42.mlp.experts.417.down_proj",
|
| 565 |
+
"model.layers.42.mlp.experts.418.down_proj",
|
| 566 |
+
"model.layers.42.mlp.experts.419.down_proj",
|
| 567 |
+
"model.layers.42.mlp.experts.42.down_proj",
|
| 568 |
+
"model.layers.42.mlp.experts.420.down_proj",
|
| 569 |
+
"model.layers.42.mlp.experts.421.down_proj",
|
| 570 |
+
"model.layers.42.mlp.experts.422.down_proj",
|
| 571 |
+
"model.layers.42.mlp.experts.423.down_proj",
|
| 572 |
+
"model.layers.42.mlp.experts.424.down_proj",
|
| 573 |
+
"model.layers.42.mlp.experts.425.down_proj",
|
| 574 |
+
"model.layers.42.mlp.experts.426.down_proj",
|
| 575 |
+
"model.layers.42.mlp.experts.427.down_proj",
|
| 576 |
+
"model.layers.42.mlp.experts.428.down_proj",
|
| 577 |
+
"model.layers.42.mlp.experts.429.down_proj",
|
| 578 |
+
"model.layers.42.mlp.experts.43.down_proj",
|
| 579 |
+
"model.layers.42.mlp.experts.430.down_proj",
|
| 580 |
+
"model.layers.42.mlp.experts.431.down_proj",
|
| 581 |
+
"model.layers.42.mlp.experts.432.down_proj",
|
| 582 |
+
"model.layers.42.mlp.experts.433.down_proj",
|
| 583 |
+
"model.layers.42.mlp.experts.434.down_proj",
|
| 584 |
+
"model.layers.42.mlp.experts.435.down_proj",
|
| 585 |
+
"model.layers.42.mlp.experts.436.down_proj",
|
| 586 |
+
"model.layers.42.mlp.experts.437.down_proj",
|
| 587 |
+
"model.layers.42.mlp.experts.438.down_proj",
|
| 588 |
+
"model.layers.42.mlp.experts.439.down_proj",
|
| 589 |
+
"model.layers.42.mlp.experts.44.down_proj",
|
| 590 |
+
"model.layers.42.mlp.experts.440.down_proj",
|
| 591 |
+
"model.layers.42.mlp.experts.441.down_proj",
|
| 592 |
+
"model.layers.42.mlp.experts.442.down_proj",
|
| 593 |
+
"model.layers.42.mlp.experts.443.down_proj",
|
| 594 |
+
"model.layers.42.mlp.experts.444.down_proj",
|
| 595 |
+
"model.layers.42.mlp.experts.445.down_proj",
|
| 596 |
+
"model.layers.42.mlp.experts.446.down_proj",
|
| 597 |
+
"model.layers.42.mlp.experts.447.down_proj",
|
| 598 |
+
"model.layers.42.mlp.experts.448.down_proj",
|
| 599 |
+
"model.layers.42.mlp.experts.449.down_proj",
|
| 600 |
+
"model.layers.42.mlp.experts.45.down_proj",
|
| 601 |
+
"model.layers.42.mlp.experts.450.down_proj",
|
| 602 |
+
"model.layers.42.mlp.experts.451.down_proj",
|
| 603 |
+
"model.layers.42.mlp.experts.452.down_proj",
|
| 604 |
+
"model.layers.42.mlp.experts.453.down_proj",
|
| 605 |
+
"model.layers.42.mlp.experts.454.down_proj",
|
| 606 |
+
"model.layers.42.mlp.experts.455.down_proj",
|
| 607 |
+
"model.layers.42.mlp.experts.456.down_proj",
|
| 608 |
+
"model.layers.42.mlp.experts.457.down_proj",
|
| 609 |
+
"model.layers.42.mlp.experts.458.down_proj",
|
| 610 |
+
"model.layers.42.mlp.experts.459.down_proj",
|
| 611 |
+
"model.layers.42.mlp.experts.46.down_proj",
|
| 612 |
+
"model.layers.42.mlp.experts.460.down_proj",
|
| 613 |
+
"model.layers.42.mlp.experts.461.down_proj",
|
| 614 |
+
"model.layers.42.mlp.experts.462.down_proj",
|
| 615 |
+
"model.layers.42.mlp.experts.463.down_proj",
|
| 616 |
+
"model.layers.42.mlp.experts.464.down_proj",
|
| 617 |
+
"model.layers.42.mlp.experts.465.down_proj",
|
| 618 |
+
"model.layers.42.mlp.experts.466.down_proj",
|
| 619 |
+
"model.layers.42.mlp.experts.467.down_proj",
|
| 620 |
+
"model.layers.42.mlp.experts.468.down_proj",
|
| 621 |
+
"model.layers.42.mlp.experts.469.down_proj",
|
| 622 |
+
"model.layers.42.mlp.experts.47.down_proj",
|
| 623 |
+
"model.layers.42.mlp.experts.470.down_proj",
|
| 624 |
+
"model.layers.42.mlp.experts.471.down_proj",
|
| 625 |
+
"model.layers.42.mlp.experts.472.down_proj",
|
| 626 |
+
"model.layers.42.mlp.experts.473.down_proj",
|
| 627 |
+
"model.layers.42.mlp.experts.474.down_proj",
|
| 628 |
+
"model.layers.42.mlp.experts.475.down_proj",
|
| 629 |
+
"model.layers.42.mlp.experts.476.down_proj",
|
| 630 |
+
"model.layers.42.mlp.experts.477.down_proj",
|
| 631 |
+
"model.layers.42.mlp.experts.478.down_proj",
|
| 632 |
+
"model.layers.42.mlp.experts.479.down_proj",
|
| 633 |
+
"model.layers.42.mlp.experts.48.down_proj",
|
| 634 |
+
"model.layers.42.mlp.experts.480.down_proj",
|
| 635 |
+
"model.layers.42.mlp.experts.481.down_proj",
|
| 636 |
+
"model.layers.42.mlp.experts.482.down_proj",
|
| 637 |
+
"model.layers.42.mlp.experts.483.down_proj",
|
| 638 |
+
"model.layers.42.mlp.experts.484.down_proj",
|
| 639 |
+
"model.layers.42.mlp.experts.485.down_proj",
|
| 640 |
+
"model.layers.42.mlp.experts.486.down_proj",
|
| 641 |
+
"model.layers.42.mlp.experts.487.down_proj",
|
| 642 |
+
"model.layers.42.mlp.experts.488.down_proj",
|
| 643 |
+
"model.layers.42.mlp.experts.489.down_proj",
|
| 644 |
+
"model.layers.42.mlp.experts.49.down_proj",
|
| 645 |
+
"model.layers.42.mlp.experts.490.down_proj",
|
| 646 |
+
"model.layers.42.mlp.experts.491.down_proj",
|
| 647 |
+
"model.layers.42.mlp.experts.492.down_proj",
|
| 648 |
+
"model.layers.42.mlp.experts.493.down_proj",
|
| 649 |
+
"model.layers.42.mlp.experts.494.down_proj",
|
| 650 |
+
"model.layers.42.mlp.experts.495.down_proj",
|
| 651 |
+
"model.layers.42.mlp.experts.496.down_proj",
|
| 652 |
+
"model.layers.42.mlp.experts.497.down_proj",
|
| 653 |
+
"model.layers.42.mlp.experts.498.down_proj",
|
| 654 |
+
"model.layers.42.mlp.experts.499.down_proj",
|
| 655 |
+
"model.layers.42.mlp.experts.5.down_proj",
|
| 656 |
+
"model.layers.42.mlp.experts.50.down_proj",
|
| 657 |
+
"model.layers.42.mlp.experts.500.down_proj",
|
| 658 |
+
"model.layers.42.mlp.experts.501.down_proj",
|
| 659 |
+
"model.layers.42.mlp.experts.502.down_proj",
|
| 660 |
+
"model.layers.42.mlp.experts.503.down_proj",
|
| 661 |
+
"model.layers.42.mlp.experts.504.down_proj",
|
| 662 |
+
"model.layers.42.mlp.experts.505.down_proj",
|
| 663 |
+
"model.layers.42.mlp.experts.506.down_proj",
|
| 664 |
+
"model.layers.42.mlp.experts.507.down_proj",
|
| 665 |
+
"model.layers.42.mlp.experts.508.down_proj",
|
| 666 |
+
"model.layers.42.mlp.experts.509.down_proj",
|
| 667 |
+
"model.layers.42.mlp.experts.51.down_proj",
|
| 668 |
+
"model.layers.42.mlp.experts.510.down_proj",
|
| 669 |
+
"model.layers.42.mlp.experts.511.down_proj",
|
| 670 |
+
"model.layers.42.mlp.experts.52.down_proj",
|
| 671 |
+
"model.layers.42.mlp.experts.53.down_proj",
|
| 672 |
+
"model.layers.42.mlp.experts.54.down_proj",
|
| 673 |
+
"model.layers.42.mlp.experts.55.down_proj",
|
| 674 |
+
"model.layers.42.mlp.experts.56.down_proj",
|
| 675 |
+
"model.layers.42.mlp.experts.57.down_proj",
|
| 676 |
+
"model.layers.42.mlp.experts.58.down_proj",
|
| 677 |
+
"model.layers.42.mlp.experts.59.down_proj",
|
| 678 |
+
"model.layers.42.mlp.experts.6.down_proj",
|
| 679 |
+
"model.layers.42.mlp.experts.60.down_proj",
|
| 680 |
+
"model.layers.42.mlp.experts.61.down_proj",
|
| 681 |
+
"model.layers.42.mlp.experts.62.down_proj",
|
| 682 |
+
"model.layers.42.mlp.experts.63.down_proj",
|
| 683 |
+
"model.layers.42.mlp.experts.64.down_proj",
|
| 684 |
+
"model.layers.42.mlp.experts.65.down_proj",
|
| 685 |
+
"model.layers.42.mlp.experts.66.down_proj",
|
| 686 |
+
"model.layers.42.mlp.experts.67.down_proj",
|
| 687 |
+
"model.layers.42.mlp.experts.68.down_proj",
|
| 688 |
+
"model.layers.42.mlp.experts.69.down_proj",
|
| 689 |
+
"model.layers.42.mlp.experts.7.down_proj",
|
| 690 |
+
"model.layers.42.mlp.experts.70.down_proj",
|
| 691 |
+
"model.layers.42.mlp.experts.71.down_proj",
|
| 692 |
+
"model.layers.42.mlp.experts.72.down_proj",
|
| 693 |
+
"model.layers.42.mlp.experts.73.down_proj",
|
| 694 |
+
"model.layers.42.mlp.experts.74.down_proj",
|
| 695 |
+
"model.layers.42.mlp.experts.75.down_proj",
|
| 696 |
+
"model.layers.42.mlp.experts.76.down_proj",
|
| 697 |
+
"model.layers.42.mlp.experts.77.down_proj",
|
| 698 |
+
"model.layers.42.mlp.experts.78.down_proj",
|
| 699 |
+
"model.layers.42.mlp.experts.79.down_proj",
|
| 700 |
+
"model.layers.42.mlp.experts.8.down_proj",
|
| 701 |
+
"model.layers.42.mlp.experts.80.down_proj",
|
| 702 |
+
"model.layers.42.mlp.experts.81.down_proj",
|
| 703 |
+
"model.layers.42.mlp.experts.82.down_proj",
|
| 704 |
+
"model.layers.42.mlp.experts.83.down_proj",
|
| 705 |
+
"model.layers.42.mlp.experts.84.down_proj",
|
| 706 |
+
"model.layers.42.mlp.experts.85.down_proj",
|
| 707 |
+
"model.layers.42.mlp.experts.86.down_proj",
|
| 708 |
+
"model.layers.42.mlp.experts.87.down_proj",
|
| 709 |
+
"model.layers.42.mlp.experts.88.down_proj",
|
| 710 |
+
"model.layers.42.mlp.experts.89.down_proj",
|
| 711 |
+
"model.layers.42.mlp.experts.9.down_proj",
|
| 712 |
+
"model.layers.42.mlp.experts.90.down_proj",
|
| 713 |
+
"model.layers.42.mlp.experts.91.down_proj",
|
| 714 |
+
"model.layers.42.mlp.experts.92.down_proj",
|
| 715 |
+
"model.layers.42.mlp.experts.93.down_proj",
|
| 716 |
+
"model.layers.42.mlp.experts.94.down_proj",
|
| 717 |
+
"model.layers.42.mlp.experts.95.down_proj",
|
| 718 |
+
"model.layers.42.mlp.experts.96.down_proj",
|
| 719 |
+
"model.layers.42.mlp.experts.97.down_proj",
|
| 720 |
+
"model.layers.42.mlp.experts.98.down_proj",
|
| 721 |
+
"model.layers.42.mlp.experts.99.down_proj",
|
| 722 |
+
"model.layers.42.mlp.experts.0.gate_proj",
|
| 723 |
+
"model.layers.42.mlp.experts.1.gate_proj",
|
| 724 |
+
"model.layers.42.mlp.experts.10.gate_proj",
|
| 725 |
+
"model.layers.42.mlp.experts.100.gate_proj",
|
| 726 |
+
"model.layers.42.mlp.experts.101.gate_proj",
|
| 727 |
+
"model.layers.42.mlp.experts.102.gate_proj",
|
| 728 |
+
"model.layers.42.mlp.experts.103.gate_proj",
|
| 729 |
+
"model.layers.42.mlp.experts.104.gate_proj",
|
| 730 |
+
"model.layers.42.mlp.experts.105.gate_proj",
|
| 731 |
+
"model.layers.42.mlp.experts.106.gate_proj",
|
| 732 |
+
"model.layers.42.mlp.experts.107.gate_proj",
|
| 733 |
+
"model.layers.42.mlp.experts.108.gate_proj",
|
| 734 |
+
"model.layers.42.mlp.experts.109.gate_proj",
|
| 735 |
+
"model.layers.42.mlp.experts.11.gate_proj",
|
| 736 |
+
"model.layers.42.mlp.experts.110.gate_proj",
|
| 737 |
+
"model.layers.42.mlp.experts.111.gate_proj",
|
| 738 |
+
"model.layers.42.mlp.experts.112.gate_proj",
|
| 739 |
+
"model.layers.42.mlp.experts.113.gate_proj",
|
| 740 |
+
"model.layers.42.mlp.experts.114.gate_proj",
|
| 741 |
+
"model.layers.42.mlp.experts.115.gate_proj",
|
| 742 |
+
"model.layers.42.mlp.experts.116.gate_proj",
|
| 743 |
+
"model.layers.42.mlp.experts.117.gate_proj",
|
| 744 |
+
"model.layers.42.mlp.experts.118.gate_proj",
|
| 745 |
+
"model.layers.42.mlp.experts.119.gate_proj",
|
| 746 |
+
"model.layers.42.mlp.experts.12.gate_proj",
|
| 747 |
+
"model.layers.42.mlp.experts.120.gate_proj",
|
| 748 |
+
"model.layers.42.mlp.experts.121.gate_proj",
|
| 749 |
+
"model.layers.42.mlp.experts.122.gate_proj",
|
| 750 |
+
"model.layers.42.mlp.experts.123.gate_proj",
|
| 751 |
+
"model.layers.42.mlp.experts.124.gate_proj",
|
| 752 |
+
"model.layers.42.mlp.experts.125.gate_proj",
|
| 753 |
+
"model.layers.42.mlp.experts.126.gate_proj",
|
| 754 |
+
"model.layers.42.mlp.experts.127.gate_proj",
|
| 755 |
+
"model.layers.42.mlp.experts.128.gate_proj",
|
| 756 |
+
"model.layers.42.mlp.experts.129.gate_proj",
|
| 757 |
+
"model.layers.42.mlp.experts.13.gate_proj",
|
| 758 |
+
"model.layers.42.mlp.experts.130.gate_proj",
|
| 759 |
+
"model.layers.42.mlp.experts.131.gate_proj",
|
| 760 |
+
"model.layers.42.mlp.experts.132.gate_proj",
|
| 761 |
+
"model.layers.42.mlp.experts.133.gate_proj",
|
| 762 |
+
"model.layers.42.mlp.experts.134.gate_proj",
|
| 763 |
+
"model.layers.42.mlp.experts.135.gate_proj",
|
| 764 |
+
"model.layers.42.mlp.experts.136.gate_proj",
|
| 765 |
+
"model.layers.42.mlp.experts.137.gate_proj",
|
| 766 |
+
"model.layers.42.mlp.experts.138.gate_proj",
|
| 767 |
+
"model.layers.42.mlp.experts.139.gate_proj",
|
| 768 |
+
"model.layers.42.mlp.experts.14.gate_proj",
|
| 769 |
+
"model.layers.42.mlp.experts.140.gate_proj",
|
| 770 |
+
"model.layers.42.mlp.experts.141.gate_proj",
|
| 771 |
+
"model.layers.42.mlp.experts.142.gate_proj",
|
| 772 |
+
"model.layers.42.mlp.experts.143.gate_proj",
|
| 773 |
+
"model.layers.42.mlp.experts.144.gate_proj",
|
| 774 |
+
"model.layers.42.mlp.experts.145.gate_proj",
|
| 775 |
+
"model.layers.42.mlp.experts.146.gate_proj",
|
| 776 |
+
"model.layers.42.mlp.experts.147.gate_proj",
|
| 777 |
+
"model.layers.42.mlp.experts.148.gate_proj",
|
| 778 |
+
"model.layers.42.mlp.experts.149.gate_proj",
|
| 779 |
+
"model.layers.42.mlp.experts.15.gate_proj",
|
| 780 |
+
"model.layers.42.mlp.experts.150.gate_proj",
|
| 781 |
+
"model.layers.42.mlp.experts.151.gate_proj",
|
| 782 |
+
"model.layers.42.mlp.experts.152.gate_proj",
|
| 783 |
+
"model.layers.42.mlp.experts.153.gate_proj",
|
| 784 |
+
"model.layers.42.mlp.experts.154.gate_proj",
|
| 785 |
+
"model.layers.42.mlp.experts.155.gate_proj",
|
| 786 |
+
"model.layers.42.mlp.experts.156.gate_proj",
|
| 787 |
+
"model.layers.42.mlp.experts.157.gate_proj",
|
| 788 |
+
"model.layers.42.mlp.experts.158.gate_proj",
|
| 789 |
+
"model.layers.42.mlp.experts.159.gate_proj",
|
| 790 |
+
"model.layers.42.mlp.experts.16.gate_proj",
|
| 791 |
+
"model.layers.42.mlp.experts.160.gate_proj",
|
| 792 |
+
"model.layers.42.mlp.experts.161.gate_proj",
|
| 793 |
+
"model.layers.42.mlp.experts.162.gate_proj",
|
| 794 |
+
"model.layers.42.mlp.experts.163.gate_proj",
|
| 795 |
+
"model.layers.42.mlp.experts.164.gate_proj",
|
| 796 |
+
"model.layers.42.mlp.experts.165.gate_proj",
|
| 797 |
+
"model.layers.42.mlp.experts.166.gate_proj",
|
| 798 |
+
"model.layers.42.mlp.experts.167.gate_proj",
|
| 799 |
+
"model.layers.42.mlp.experts.168.gate_proj",
|
| 800 |
+
"model.layers.42.mlp.experts.169.gate_proj",
|
| 801 |
+
"model.layers.42.mlp.experts.17.gate_proj",
|
| 802 |
+
"model.layers.42.mlp.experts.170.gate_proj",
|
| 803 |
+
"model.layers.42.mlp.experts.171.gate_proj",
|
| 804 |
+
"model.layers.42.mlp.experts.172.gate_proj",
|
| 805 |
+
"model.layers.42.mlp.experts.173.gate_proj",
|
| 806 |
+
"model.layers.42.mlp.experts.174.gate_proj",
|
| 807 |
+
"model.layers.42.mlp.experts.175.gate_proj",
|
| 808 |
+
"model.layers.42.mlp.experts.176.gate_proj",
|
| 809 |
+
"model.layers.42.mlp.experts.177.gate_proj",
|
| 810 |
+
"model.layers.42.mlp.experts.178.gate_proj",
|
| 811 |
+
"model.layers.42.mlp.experts.179.gate_proj",
|
| 812 |
+
"model.layers.42.mlp.experts.18.gate_proj",
|
| 813 |
+
"model.layers.42.mlp.experts.180.gate_proj",
|
| 814 |
+
"model.layers.42.mlp.experts.181.gate_proj",
|
| 815 |
+
"model.layers.42.mlp.experts.182.gate_proj",
|
| 816 |
+
"model.layers.42.mlp.experts.183.gate_proj",
|
| 817 |
+
"model.layers.42.mlp.experts.184.gate_proj",
|
| 818 |
+
"model.layers.42.mlp.experts.185.gate_proj",
|
| 819 |
+
"model.layers.42.mlp.experts.186.gate_proj",
|
| 820 |
+
"model.layers.42.mlp.experts.187.gate_proj",
|
| 821 |
+
"model.layers.42.mlp.experts.188.gate_proj",
|
| 822 |
+
"model.layers.42.mlp.experts.189.gate_proj",
|
| 823 |
+
"model.layers.42.mlp.experts.19.gate_proj",
|
| 824 |
+
"model.layers.42.mlp.experts.190.gate_proj",
|
| 825 |
+
"model.layers.42.mlp.experts.191.gate_proj",
|
| 826 |
+
"model.layers.42.mlp.experts.192.gate_proj",
|
| 827 |
+
"model.layers.42.mlp.experts.193.gate_proj",
|
| 828 |
+
"model.layers.42.mlp.experts.194.gate_proj",
|
| 829 |
+
"model.layers.42.mlp.experts.195.gate_proj",
|
| 830 |
+
"model.layers.42.mlp.experts.196.gate_proj",
|
| 831 |
+
"model.layers.42.mlp.experts.197.gate_proj",
|
| 832 |
+
"model.layers.42.mlp.experts.198.gate_proj",
|
| 833 |
+
"model.layers.42.mlp.experts.199.gate_proj",
|
| 834 |
+
"model.layers.42.mlp.experts.2.gate_proj",
|
| 835 |
+
"model.layers.42.mlp.experts.20.gate_proj",
|
| 836 |
+
"model.layers.42.mlp.experts.200.gate_proj",
|
| 837 |
+
"model.layers.42.mlp.experts.201.gate_proj",
|
| 838 |
+
"model.layers.42.mlp.experts.202.gate_proj",
|
| 839 |
+
"model.layers.42.mlp.experts.203.gate_proj",
|
| 840 |
+
"model.layers.42.mlp.experts.204.gate_proj",
|
| 841 |
+
"model.layers.42.mlp.experts.205.gate_proj",
|
| 842 |
+
"model.layers.42.mlp.experts.206.gate_proj",
|
| 843 |
+
"model.layers.42.mlp.experts.207.gate_proj",
|
| 844 |
+
"model.layers.42.mlp.experts.208.gate_proj",
|
| 845 |
+
"model.layers.42.mlp.experts.209.gate_proj",
|
| 846 |
+
"model.layers.42.mlp.experts.21.gate_proj",
|
| 847 |
+
"model.layers.42.mlp.experts.210.gate_proj",
|
| 848 |
+
"model.layers.42.mlp.experts.211.gate_proj",
|
| 849 |
+
"model.layers.42.mlp.experts.212.gate_proj",
|
| 850 |
+
"model.layers.42.mlp.experts.213.gate_proj",
|
| 851 |
+
"model.layers.42.mlp.experts.214.gate_proj",
|
| 852 |
+
"model.layers.42.mlp.experts.215.gate_proj",
|
| 853 |
+
"model.layers.42.mlp.experts.216.gate_proj",
|
| 854 |
+
"model.layers.42.mlp.experts.217.gate_proj",
|
| 855 |
+
"model.layers.42.mlp.experts.218.gate_proj",
|
| 856 |
+
"model.layers.42.mlp.experts.219.gate_proj",
|
| 857 |
+
"model.layers.42.mlp.experts.22.gate_proj",
|
| 858 |
+
"model.layers.42.mlp.experts.220.gate_proj",
|
| 859 |
+
"model.layers.42.mlp.experts.221.gate_proj",
|
| 860 |
+
"model.layers.42.mlp.experts.222.gate_proj",
|
| 861 |
+
"model.layers.42.mlp.experts.223.gate_proj",
|
| 862 |
+
"model.layers.42.mlp.experts.224.gate_proj",
|
| 863 |
+
"model.layers.42.mlp.experts.225.gate_proj",
|
| 864 |
+
"model.layers.42.mlp.experts.226.gate_proj",
|
| 865 |
+
"model.layers.42.mlp.experts.227.gate_proj",
|
| 866 |
+
"model.layers.42.mlp.experts.228.gate_proj",
|
| 867 |
+
"model.layers.42.mlp.experts.229.gate_proj",
|
| 868 |
+
"model.layers.42.mlp.experts.23.gate_proj",
|
| 869 |
+
"model.layers.42.mlp.experts.230.gate_proj",
|
| 870 |
+
"model.layers.42.mlp.experts.231.gate_proj",
|
| 871 |
+
"model.layers.42.mlp.experts.232.gate_proj",
|
| 872 |
+
"model.layers.42.mlp.experts.233.gate_proj",
|
| 873 |
+
"model.layers.42.mlp.experts.234.gate_proj",
|
| 874 |
+
"model.layers.42.mlp.experts.235.gate_proj",
|
| 875 |
+
"model.layers.42.mlp.experts.236.gate_proj",
|
| 876 |
+
"model.layers.42.mlp.experts.237.gate_proj",
|
| 877 |
+
"model.layers.42.mlp.experts.238.gate_proj",
|
| 878 |
+
"model.layers.42.mlp.experts.239.gate_proj",
|
| 879 |
+
"model.layers.42.mlp.experts.24.gate_proj",
|
| 880 |
+
"model.layers.42.mlp.experts.240.gate_proj",
|
| 881 |
+
"model.layers.42.mlp.experts.241.gate_proj",
|
| 882 |
+
"model.layers.42.mlp.experts.242.gate_proj",
|
| 883 |
+
"model.layers.42.mlp.experts.243.gate_proj",
|
| 884 |
+
"model.layers.42.mlp.experts.244.gate_proj",
|
| 885 |
+
"model.layers.42.mlp.experts.245.gate_proj",
|
| 886 |
+
"model.layers.42.mlp.experts.246.gate_proj",
|
| 887 |
+
"model.layers.42.mlp.experts.247.gate_proj",
|
| 888 |
+
"model.layers.42.mlp.experts.248.gate_proj",
|
| 889 |
+
"model.layers.42.mlp.experts.249.gate_proj",
|
| 890 |
+
"model.layers.42.mlp.experts.25.gate_proj",
|
| 891 |
+
"model.layers.42.mlp.experts.250.gate_proj",
|
| 892 |
+
"model.layers.42.mlp.experts.251.gate_proj",
|
| 893 |
+
"model.layers.42.mlp.experts.252.gate_proj",
|
| 894 |
+
"model.layers.42.mlp.experts.253.gate_proj",
|
| 895 |
+
"model.layers.42.mlp.experts.254.gate_proj",
|
| 896 |
+
"model.layers.42.mlp.experts.255.gate_proj",
|
| 897 |
+
"model.layers.42.mlp.experts.256.gate_proj",
|
| 898 |
+
"model.layers.42.mlp.experts.257.gate_proj",
|
| 899 |
+
"model.layers.42.mlp.experts.258.gate_proj",
|
| 900 |
+
"model.layers.42.mlp.experts.259.gate_proj",
|
| 901 |
+
"model.layers.42.mlp.experts.26.gate_proj",
|
| 902 |
+
"model.layers.42.mlp.experts.260.gate_proj",
|
| 903 |
+
"model.layers.42.mlp.experts.261.gate_proj",
|
| 904 |
+
"model.layers.42.mlp.experts.262.gate_proj",
|
| 905 |
+
"model.layers.42.mlp.experts.263.gate_proj",
|
| 906 |
+
"model.layers.42.mlp.experts.264.gate_proj",
|
| 907 |
+
"model.layers.42.mlp.experts.265.gate_proj",
|
| 908 |
+
"model.layers.42.mlp.experts.266.gate_proj",
|
| 909 |
+
"model.layers.42.mlp.experts.267.gate_proj",
|
| 910 |
+
"model.layers.42.mlp.experts.268.gate_proj",
|
| 911 |
+
"model.layers.42.mlp.experts.269.gate_proj",
|
| 912 |
+
"model.layers.42.mlp.experts.27.gate_proj",
|
| 913 |
+
"model.layers.42.mlp.experts.270.gate_proj",
|
| 914 |
+
"model.layers.42.mlp.experts.271.gate_proj",
|
| 915 |
+
"model.layers.42.mlp.experts.272.gate_proj",
|
| 916 |
+
"model.layers.42.mlp.experts.273.gate_proj",
|
| 917 |
+
"model.layers.42.mlp.experts.274.gate_proj",
|
| 918 |
+
"model.layers.42.mlp.experts.275.gate_proj",
|
| 919 |
+
"model.layers.42.mlp.experts.276.gate_proj",
|
| 920 |
+
"model.layers.42.mlp.experts.277.gate_proj",
|
| 921 |
+
"model.layers.42.mlp.experts.278.gate_proj",
|
| 922 |
+
"model.layers.42.mlp.experts.279.gate_proj",
|
| 923 |
+
"model.layers.42.mlp.experts.28.gate_proj",
|
| 924 |
+
"model.layers.42.mlp.experts.280.gate_proj",
|
| 925 |
+
"model.layers.42.mlp.experts.281.gate_proj",
|
| 926 |
+
"model.layers.42.mlp.experts.282.gate_proj",
|
| 927 |
+
"model.layers.42.mlp.experts.283.gate_proj",
|
| 928 |
+
"model.layers.42.mlp.experts.284.gate_proj",
|
| 929 |
+
"model.layers.42.mlp.experts.285.gate_proj",
|
| 930 |
+
"model.layers.42.mlp.experts.286.gate_proj",
|
| 931 |
+
"model.layers.42.mlp.experts.287.gate_proj",
|
| 932 |
+
"model.layers.42.mlp.experts.288.gate_proj",
|
| 933 |
+
"model.layers.42.mlp.experts.289.gate_proj",
|
| 934 |
+
"model.layers.42.mlp.experts.29.gate_proj",
|
| 935 |
+
"model.layers.42.mlp.experts.290.gate_proj",
|
| 936 |
+
"model.layers.42.mlp.experts.291.gate_proj",
|
| 937 |
+
"model.layers.42.mlp.experts.292.gate_proj",
|
| 938 |
+
"model.layers.42.mlp.experts.293.gate_proj",
|
| 939 |
+
"model.layers.42.mlp.experts.294.gate_proj",
|
| 940 |
+
"model.layers.42.mlp.experts.295.gate_proj",
|
| 941 |
+
"model.layers.42.mlp.experts.296.gate_proj",
|
| 942 |
+
"model.layers.42.mlp.experts.297.gate_proj",
|
| 943 |
+
"model.layers.42.mlp.experts.298.gate_proj",
|
| 944 |
+
"model.layers.42.mlp.experts.299.gate_proj",
|
| 945 |
+
"model.layers.42.mlp.experts.3.gate_proj",
|
| 946 |
+
"model.layers.42.mlp.experts.30.gate_proj",
|
| 947 |
+
"model.layers.42.mlp.experts.300.gate_proj",
|
| 948 |
+
"model.layers.42.mlp.experts.301.gate_proj",
|
| 949 |
+
"model.layers.42.mlp.experts.302.gate_proj",
|
| 950 |
+
"model.layers.42.mlp.experts.303.gate_proj",
|
| 951 |
+
"model.layers.42.mlp.experts.304.gate_proj",
|
| 952 |
+
"model.layers.42.mlp.experts.305.gate_proj",
|
| 953 |
+
"model.layers.42.mlp.experts.306.gate_proj",
|
| 954 |
+
"model.layers.42.mlp.experts.307.gate_proj",
|
| 955 |
+
"model.layers.42.mlp.experts.308.gate_proj",
|
| 956 |
+
"model.layers.42.mlp.experts.309.gate_proj",
|
| 957 |
+
"model.layers.42.mlp.experts.31.gate_proj",
|
| 958 |
+
"model.layers.42.mlp.experts.310.gate_proj",
|
| 959 |
+
"model.layers.42.mlp.experts.311.gate_proj",
|
| 960 |
+
"model.layers.42.mlp.experts.312.gate_proj",
|
| 961 |
+
"model.layers.42.mlp.experts.313.gate_proj",
|
| 962 |
+
"model.layers.42.mlp.experts.314.gate_proj",
|
| 963 |
+
"model.layers.42.mlp.experts.315.gate_proj",
|
| 964 |
+
"model.layers.42.mlp.experts.316.gate_proj",
|
| 965 |
+
"model.layers.42.mlp.experts.317.gate_proj",
|
| 966 |
+
"model.layers.42.mlp.experts.318.gate_proj",
|
| 967 |
+
"model.layers.42.mlp.experts.319.gate_proj",
|
| 968 |
+
"model.layers.42.mlp.experts.32.gate_proj",
|
| 969 |
+
"model.layers.42.mlp.experts.320.gate_proj",
|
| 970 |
+
"model.layers.42.mlp.experts.321.gate_proj",
|
| 971 |
+
"model.layers.42.mlp.experts.322.gate_proj",
|
| 972 |
+
"model.layers.42.mlp.experts.323.gate_proj",
|
| 973 |
+
"model.layers.42.mlp.experts.324.gate_proj",
|
| 974 |
+
"model.layers.42.mlp.experts.325.gate_proj",
|
| 975 |
+
"model.layers.42.mlp.experts.326.gate_proj",
|
| 976 |
+
"model.layers.42.mlp.experts.327.gate_proj",
|
| 977 |
+
"model.layers.42.mlp.experts.328.gate_proj",
|
| 978 |
+
"model.layers.42.mlp.experts.329.gate_proj",
|
| 979 |
+
"model.layers.42.mlp.experts.33.gate_proj",
|
| 980 |
+
"model.layers.42.mlp.experts.330.gate_proj",
|
| 981 |
+
"model.layers.42.mlp.experts.331.gate_proj",
|
| 982 |
+
"model.layers.42.mlp.experts.332.gate_proj",
|
| 983 |
+
"model.layers.42.mlp.experts.333.gate_proj",
|
| 984 |
+
"model.layers.42.mlp.experts.334.gate_proj",
|
| 985 |
+
"model.layers.42.mlp.experts.335.gate_proj",
|
| 986 |
+
"model.layers.42.mlp.experts.336.gate_proj",
|
| 987 |
+
"model.layers.42.mlp.experts.337.gate_proj",
|
| 988 |
+
"model.layers.42.mlp.experts.338.gate_proj",
|
| 989 |
+
"model.layers.42.mlp.experts.339.gate_proj",
|
| 990 |
+
"model.layers.42.mlp.experts.34.gate_proj",
|
| 991 |
+
"model.layers.42.mlp.experts.340.gate_proj",
|
| 992 |
+
"model.layers.42.mlp.experts.341.gate_proj",
|
| 993 |
+
"model.layers.42.mlp.experts.342.gate_proj",
|
| 994 |
+
"model.layers.42.mlp.experts.343.gate_proj",
|
| 995 |
+
"model.layers.42.mlp.experts.344.gate_proj",
|
| 996 |
+
"model.layers.42.mlp.experts.345.gate_proj",
|
| 997 |
+
"model.layers.42.mlp.experts.346.gate_proj",
|
| 998 |
+
"model.layers.42.mlp.experts.347.gate_proj",
|
| 999 |
+
"model.layers.42.mlp.experts.348.gate_proj",
|
| 1000 |
+
"model.layers.42.mlp.experts.349.gate_proj",
|
| 1001 |
+
"model.layers.42.mlp.experts.35.gate_proj",
|
| 1002 |
+
"model.layers.42.mlp.experts.350.gate_proj",
|
| 1003 |
+
"model.layers.42.mlp.experts.351.gate_proj",
|
| 1004 |
+
"model.layers.42.mlp.experts.352.gate_proj",
|
| 1005 |
+
"model.layers.42.mlp.experts.353.gate_proj",
|
| 1006 |
+
"model.layers.42.mlp.experts.354.gate_proj",
|
| 1007 |
+
"model.layers.42.mlp.experts.355.gate_proj",
|
| 1008 |
+
"model.layers.42.mlp.experts.356.gate_proj",
|
| 1009 |
+
"model.layers.42.mlp.experts.357.gate_proj",
|
| 1010 |
+
"model.layers.42.mlp.experts.358.gate_proj",
|
| 1011 |
+
"model.layers.42.mlp.experts.359.gate_proj",
|
| 1012 |
+
"model.layers.42.mlp.experts.36.gate_proj",
|
| 1013 |
+
"model.layers.42.mlp.experts.360.gate_proj",
|
| 1014 |
+
"model.layers.42.mlp.experts.361.gate_proj",
|
| 1015 |
+
"model.layers.42.mlp.experts.362.gate_proj",
|
| 1016 |
+
"model.layers.42.mlp.experts.363.gate_proj",
|
| 1017 |
+
"model.layers.42.mlp.experts.364.gate_proj",
|
| 1018 |
+
"model.layers.42.mlp.experts.365.gate_proj",
|
| 1019 |
+
"model.layers.42.mlp.experts.366.gate_proj",
|
| 1020 |
+
"model.layers.42.mlp.experts.367.gate_proj",
|
| 1021 |
+
"model.layers.42.mlp.experts.368.gate_proj",
|
| 1022 |
+
"model.layers.42.mlp.experts.369.gate_proj",
|
| 1023 |
+
"model.layers.42.mlp.experts.37.gate_proj",
|
| 1024 |
+
"model.layers.42.mlp.experts.370.gate_proj",
|
| 1025 |
+
"model.layers.42.mlp.experts.371.gate_proj",
|
| 1026 |
+
"model.layers.42.mlp.experts.372.gate_proj",
|
| 1027 |
+
"model.layers.42.mlp.experts.373.gate_proj",
|
| 1028 |
+
"model.layers.42.mlp.experts.374.gate_proj",
|
| 1029 |
+
"model.layers.42.mlp.experts.375.gate_proj",
|
| 1030 |
+
"model.layers.42.mlp.experts.376.gate_proj",
|
| 1031 |
+
"model.layers.42.mlp.experts.377.gate_proj",
|
| 1032 |
+
"model.layers.42.mlp.experts.378.gate_proj",
|
| 1033 |
+
"model.layers.42.mlp.experts.379.gate_proj",
|
| 1034 |
+
"model.layers.42.mlp.experts.38.gate_proj",
|
| 1035 |
+
"model.layers.42.mlp.experts.380.gate_proj",
|
| 1036 |
+
"model.layers.42.mlp.experts.381.gate_proj",
|
| 1037 |
+
"model.layers.42.mlp.experts.382.gate_proj",
|
| 1038 |
+
"model.layers.42.mlp.experts.383.gate_proj",
|
| 1039 |
+
"model.layers.42.mlp.experts.384.gate_proj",
|
| 1040 |
+
"model.layers.42.mlp.experts.385.gate_proj",
|
| 1041 |
+
"model.layers.42.mlp.experts.386.gate_proj",
|
| 1042 |
+
"model.layers.42.mlp.experts.387.gate_proj",
|
| 1043 |
+
"model.layers.42.mlp.experts.388.gate_proj",
|
| 1044 |
+
"model.layers.42.mlp.experts.389.gate_proj",
|
| 1045 |
+
"model.layers.42.mlp.experts.39.gate_proj",
|
| 1046 |
+
"model.layers.42.mlp.experts.390.gate_proj",
|
| 1047 |
+
"model.layers.42.mlp.experts.391.gate_proj",
|
| 1048 |
+
"model.layers.42.mlp.experts.392.gate_proj",
|
| 1049 |
+
"model.layers.42.mlp.experts.393.gate_proj",
|
| 1050 |
+
"model.layers.42.mlp.experts.394.gate_proj",
|
| 1051 |
+
"model.layers.42.mlp.experts.395.gate_proj",
|
| 1052 |
+
"model.layers.42.mlp.experts.396.gate_proj",
|
| 1053 |
+
"model.layers.42.mlp.experts.397.gate_proj",
|
| 1054 |
+
"model.layers.42.mlp.experts.398.gate_proj",
|
| 1055 |
+
"model.layers.42.mlp.experts.399.gate_proj",
|
| 1056 |
+
"model.layers.42.mlp.experts.4.gate_proj",
|
| 1057 |
+
"model.layers.42.mlp.experts.40.gate_proj",
|
| 1058 |
+
"model.layers.42.mlp.experts.400.gate_proj",
|
| 1059 |
+
"model.layers.42.mlp.experts.401.gate_proj",
|
| 1060 |
+
"model.layers.42.mlp.experts.402.gate_proj",
|
| 1061 |
+
"model.layers.42.mlp.experts.403.gate_proj",
|
| 1062 |
+
"model.layers.42.mlp.experts.404.gate_proj",
|
| 1063 |
+
"model.layers.42.mlp.experts.405.gate_proj",
|
| 1064 |
+
"model.layers.42.mlp.experts.406.gate_proj",
|
| 1065 |
+
"model.layers.42.mlp.experts.407.gate_proj",
|
| 1066 |
+
"model.layers.42.mlp.experts.408.gate_proj",
|
| 1067 |
+
"model.layers.42.mlp.experts.409.gate_proj",
|
| 1068 |
+
"model.layers.42.mlp.experts.41.gate_proj",
|
| 1069 |
+
"model.layers.42.mlp.experts.410.gate_proj",
|
| 1070 |
+
"model.layers.42.mlp.experts.411.gate_proj",
|
| 1071 |
+
"model.layers.42.mlp.experts.412.gate_proj",
|
| 1072 |
+
"model.layers.42.mlp.experts.413.gate_proj",
|
| 1073 |
+
"model.layers.42.mlp.experts.414.gate_proj",
|
| 1074 |
+
"model.layers.42.mlp.experts.415.gate_proj",
|
| 1075 |
+
"model.layers.42.mlp.experts.416.gate_proj",
|
| 1076 |
+
"model.layers.42.mlp.experts.417.gate_proj",
|
| 1077 |
+
"model.layers.42.mlp.experts.418.gate_proj",
|
| 1078 |
+
"model.layers.42.mlp.experts.419.gate_proj",
|
| 1079 |
+
"model.layers.42.mlp.experts.42.gate_proj",
|
| 1080 |
+
"model.layers.42.mlp.experts.420.gate_proj",
|
| 1081 |
+
"model.layers.42.mlp.experts.421.gate_proj",
|
| 1082 |
+
"model.layers.42.mlp.experts.422.gate_proj",
|
| 1083 |
+
"model.layers.42.mlp.experts.423.gate_proj",
|
| 1084 |
+
"model.layers.42.mlp.experts.424.gate_proj",
|
| 1085 |
+
"model.layers.42.mlp.experts.425.gate_proj",
|
| 1086 |
+
"model.layers.42.mlp.experts.426.gate_proj",
|
| 1087 |
+
"model.layers.42.mlp.experts.427.gate_proj",
|
| 1088 |
+
"model.layers.42.mlp.experts.428.gate_proj",
|
| 1089 |
+
"model.layers.42.mlp.experts.429.gate_proj",
|
| 1090 |
+
"model.layers.42.mlp.experts.43.gate_proj",
|
| 1091 |
+
"model.layers.42.mlp.experts.430.gate_proj",
|
| 1092 |
+
"model.layers.42.mlp.experts.431.gate_proj",
|
| 1093 |
+
"model.layers.42.mlp.experts.432.gate_proj",
|
| 1094 |
+
"model.layers.42.mlp.experts.433.gate_proj",
|
| 1095 |
+
"model.layers.42.mlp.experts.434.gate_proj",
|
| 1096 |
+
"model.layers.42.mlp.experts.435.gate_proj",
|
| 1097 |
+
"model.layers.42.mlp.experts.436.gate_proj",
|
| 1098 |
+
"model.layers.42.mlp.experts.437.gate_proj",
|
| 1099 |
+
"model.layers.42.mlp.experts.438.gate_proj",
|
| 1100 |
+
"model.layers.42.mlp.experts.439.gate_proj",
|
| 1101 |
+
"model.layers.42.mlp.experts.44.gate_proj",
|
| 1102 |
+
"model.layers.42.mlp.experts.440.gate_proj",
|
| 1103 |
+
"model.layers.42.mlp.experts.441.gate_proj",
|
| 1104 |
+
"model.layers.42.mlp.experts.442.gate_proj",
|
| 1105 |
+
"model.layers.42.mlp.experts.443.gate_proj",
|
| 1106 |
+
"model.layers.42.mlp.experts.444.gate_proj",
|
| 1107 |
+
"model.layers.42.mlp.experts.445.gate_proj",
|
| 1108 |
+
"model.layers.42.mlp.experts.446.gate_proj",
|
| 1109 |
+
"model.layers.42.mlp.experts.447.gate_proj",
|
| 1110 |
+
"model.layers.42.mlp.experts.448.gate_proj",
|
| 1111 |
+
"model.layers.42.mlp.experts.449.gate_proj",
|
| 1112 |
+
"model.layers.42.mlp.experts.45.gate_proj",
|
| 1113 |
+
"model.layers.42.mlp.experts.450.gate_proj",
|
| 1114 |
+
"model.layers.42.mlp.experts.451.gate_proj",
|
| 1115 |
+
"model.layers.42.mlp.experts.452.gate_proj",
|
| 1116 |
+
"model.layers.42.mlp.experts.453.gate_proj",
|
| 1117 |
+
"model.layers.42.mlp.experts.454.gate_proj",
|
| 1118 |
+
"model.layers.42.mlp.experts.455.gate_proj",
|
| 1119 |
+
"model.layers.42.mlp.experts.456.gate_proj",
|
| 1120 |
+
"model.layers.42.mlp.experts.457.gate_proj",
|
| 1121 |
+
"model.layers.42.mlp.experts.458.gate_proj",
|
| 1122 |
+
"model.layers.42.mlp.experts.459.gate_proj",
|
| 1123 |
+
"model.layers.42.mlp.experts.46.gate_proj",
|
| 1124 |
+
"model.layers.42.mlp.experts.460.gate_proj",
|
| 1125 |
+
"model.layers.42.mlp.experts.461.gate_proj",
|
| 1126 |
+
"model.layers.42.mlp.experts.462.gate_proj",
|
| 1127 |
+
"model.layers.42.mlp.experts.463.gate_proj",
|
| 1128 |
+
"model.layers.42.mlp.experts.464.gate_proj",
|
| 1129 |
+
"model.layers.42.mlp.experts.465.gate_proj",
|
| 1130 |
+
"model.layers.42.mlp.experts.466.gate_proj",
|
| 1131 |
+
"model.layers.42.mlp.experts.467.gate_proj",
|
| 1132 |
+
"model.layers.42.mlp.experts.468.gate_proj",
|
| 1133 |
+
"model.layers.42.mlp.experts.469.gate_proj",
|
| 1134 |
+
"model.layers.42.mlp.experts.47.gate_proj",
|
| 1135 |
+
"model.layers.42.mlp.experts.470.gate_proj",
|
| 1136 |
+
"model.layers.42.mlp.experts.471.gate_proj",
|
| 1137 |
+
"model.layers.42.mlp.experts.472.gate_proj",
|
| 1138 |
+
"model.layers.42.mlp.experts.473.gate_proj",
|
| 1139 |
+
"model.layers.42.mlp.experts.474.gate_proj",
|
| 1140 |
+
"model.layers.42.mlp.experts.475.gate_proj",
|
| 1141 |
+
"model.layers.42.mlp.experts.476.gate_proj",
|
| 1142 |
+
"model.layers.42.mlp.experts.477.gate_proj",
|
| 1143 |
+
"model.layers.42.mlp.experts.478.gate_proj",
|
| 1144 |
+
"model.layers.42.mlp.experts.479.gate_proj",
|
| 1145 |
+
"model.layers.42.mlp.experts.48.gate_proj",
|
| 1146 |
+
"model.layers.42.mlp.experts.480.gate_proj",
|
| 1147 |
+
"model.layers.42.mlp.experts.481.gate_proj",
|
| 1148 |
+
"model.layers.42.mlp.experts.482.gate_proj",
|
| 1149 |
+
"model.layers.42.mlp.experts.483.gate_proj",
|
| 1150 |
+
"model.layers.42.mlp.experts.484.gate_proj",
|
| 1151 |
+
"model.layers.42.mlp.experts.485.gate_proj",
|
| 1152 |
+
"model.layers.42.mlp.experts.486.gate_proj",
|
| 1153 |
+
"model.layers.42.mlp.experts.487.gate_proj",
|
| 1154 |
+
"model.layers.42.mlp.experts.488.gate_proj",
|
| 1155 |
+
"model.layers.42.mlp.experts.489.gate_proj",
|
| 1156 |
+
"model.layers.42.mlp.experts.49.gate_proj",
|
| 1157 |
+
"model.layers.42.mlp.experts.490.gate_proj",
|
| 1158 |
+
"model.layers.42.mlp.experts.491.gate_proj",
|
| 1159 |
+
"model.layers.42.mlp.experts.492.gate_proj",
|
| 1160 |
+
"model.layers.42.mlp.experts.493.gate_proj",
|
| 1161 |
+
"model.layers.42.mlp.experts.494.gate_proj",
|
| 1162 |
+
"model.layers.42.mlp.experts.495.gate_proj",
|
| 1163 |
+
"model.layers.42.mlp.experts.496.gate_proj",
|
| 1164 |
+
"model.layers.42.mlp.experts.497.gate_proj",
|
| 1165 |
+
"model.layers.42.mlp.experts.498.gate_proj",
|
| 1166 |
+
"model.layers.42.mlp.experts.499.gate_proj",
|
| 1167 |
+
"model.layers.42.mlp.experts.5.gate_proj",
|
| 1168 |
+
"model.layers.42.mlp.experts.50.gate_proj",
|
| 1169 |
+
"model.layers.42.mlp.experts.500.gate_proj",
|
| 1170 |
+
"model.layers.42.mlp.experts.501.gate_proj",
|
| 1171 |
+
"model.layers.42.mlp.experts.502.gate_proj",
|
| 1172 |
+
"model.layers.42.mlp.experts.503.gate_proj",
|
| 1173 |
+
"model.layers.42.mlp.experts.504.gate_proj",
|
| 1174 |
+
"model.layers.42.mlp.experts.505.gate_proj",
|
| 1175 |
+
"model.layers.42.mlp.experts.506.gate_proj",
|
| 1176 |
+
"model.layers.42.mlp.experts.507.gate_proj",
|
| 1177 |
+
"model.layers.42.mlp.experts.508.gate_proj",
|
| 1178 |
+
"model.layers.42.mlp.experts.509.gate_proj",
|
| 1179 |
+
"model.layers.42.mlp.experts.51.gate_proj",
|
| 1180 |
+
"model.layers.42.mlp.experts.510.gate_proj",
|
| 1181 |
+
"model.layers.42.mlp.experts.511.gate_proj",
|
| 1182 |
+
"model.layers.42.mlp.experts.52.gate_proj",
|
| 1183 |
+
"model.layers.42.mlp.experts.53.gate_proj",
|
| 1184 |
+
"model.layers.42.mlp.experts.54.gate_proj",
|
| 1185 |
+
"model.layers.42.mlp.experts.55.gate_proj",
|
| 1186 |
+
"model.layers.42.mlp.experts.56.gate_proj",
|
| 1187 |
+
"model.layers.42.mlp.experts.57.gate_proj",
|
| 1188 |
+
"model.layers.42.mlp.experts.58.gate_proj",
|
| 1189 |
+
"model.layers.42.mlp.experts.59.gate_proj",
|
| 1190 |
+
"model.layers.42.mlp.experts.6.gate_proj",
|
| 1191 |
+
"model.layers.42.mlp.experts.60.gate_proj",
|
| 1192 |
+
"model.layers.42.mlp.experts.61.gate_proj",
|
| 1193 |
+
"model.layers.42.mlp.experts.62.gate_proj",
|
| 1194 |
+
"model.layers.42.mlp.experts.63.gate_proj",
|
| 1195 |
+
"model.layers.42.mlp.experts.64.gate_proj",
|
| 1196 |
+
"model.layers.42.mlp.experts.65.gate_proj",
|
| 1197 |
+
"model.layers.42.mlp.experts.66.gate_proj",
|
| 1198 |
+
"model.layers.42.mlp.experts.67.gate_proj",
|
| 1199 |
+
"model.layers.42.mlp.experts.68.gate_proj",
|
| 1200 |
+
"model.layers.42.mlp.experts.69.gate_proj",
|
| 1201 |
+
"model.layers.42.mlp.experts.7.gate_proj",
|
| 1202 |
+
"model.layers.42.mlp.experts.70.gate_proj",
|
| 1203 |
+
"model.layers.42.mlp.experts.71.gate_proj",
|
| 1204 |
+
"model.layers.42.mlp.experts.72.gate_proj",
|
| 1205 |
+
"model.layers.42.mlp.experts.73.gate_proj",
|
| 1206 |
+
"model.layers.42.mlp.experts.74.gate_proj",
|
| 1207 |
+
"model.layers.42.mlp.experts.75.gate_proj",
|
| 1208 |
+
"model.layers.42.mlp.experts.76.gate_proj",
|
| 1209 |
+
"model.layers.42.mlp.experts.77.gate_proj",
|
| 1210 |
+
"model.layers.42.mlp.experts.78.gate_proj",
|
| 1211 |
+
"model.layers.42.mlp.experts.79.gate_proj",
|
| 1212 |
+
"model.layers.42.mlp.experts.8.gate_proj",
|
| 1213 |
+
"model.layers.42.mlp.experts.80.gate_proj",
|
| 1214 |
+
"model.layers.42.mlp.experts.81.gate_proj",
|
| 1215 |
+
"model.layers.42.mlp.experts.82.gate_proj",
|
| 1216 |
+
"model.layers.42.mlp.experts.83.gate_proj",
|
| 1217 |
+
"model.layers.42.mlp.experts.84.gate_proj",
|
| 1218 |
+
"model.layers.42.mlp.experts.85.gate_proj",
|
| 1219 |
+
"model.layers.42.mlp.experts.86.gate_proj",
|
| 1220 |
+
"model.layers.42.mlp.experts.87.gate_proj",
|
| 1221 |
+
"model.layers.42.mlp.experts.88.gate_proj",
|
| 1222 |
+
"model.layers.42.mlp.experts.89.gate_proj",
|
| 1223 |
+
"model.layers.42.mlp.experts.9.gate_proj",
|
| 1224 |
+
"model.layers.42.mlp.experts.90.gate_proj",
|
| 1225 |
+
"model.layers.42.mlp.experts.91.gate_proj",
|
| 1226 |
+
"model.layers.42.mlp.experts.92.gate_proj",
|
| 1227 |
+
"model.layers.42.mlp.experts.93.gate_proj",
|
| 1228 |
+
"model.layers.42.mlp.experts.94.gate_proj",
|
| 1229 |
+
"model.layers.42.mlp.experts.95.gate_proj",
|
| 1230 |
+
"model.layers.42.mlp.experts.96.gate_proj",
|
| 1231 |
+
"model.layers.42.mlp.experts.97.gate_proj",
|
| 1232 |
+
"model.layers.42.mlp.experts.98.gate_proj",
|
| 1233 |
+
"model.layers.42.mlp.experts.99.gate_proj",
|
| 1234 |
+
"model.layers.42.mlp.experts.0.up_proj",
|
| 1235 |
+
"model.layers.42.mlp.experts.1.up_proj",
|
| 1236 |
+
"model.layers.42.mlp.experts.10.up_proj",
|
| 1237 |
+
"model.layers.42.mlp.experts.100.up_proj",
|
| 1238 |
+
"model.layers.42.mlp.experts.101.up_proj",
|
| 1239 |
+
"model.layers.42.mlp.experts.102.up_proj",
|
| 1240 |
+
"model.layers.42.mlp.experts.103.up_proj",
|
| 1241 |
+
"model.layers.42.mlp.experts.104.up_proj",
|
| 1242 |
+
"model.layers.42.mlp.experts.105.up_proj",
|
| 1243 |
+
"model.layers.42.mlp.experts.106.up_proj",
|
| 1244 |
+
"model.layers.42.mlp.experts.107.up_proj",
|
| 1245 |
+
"model.layers.42.mlp.experts.108.up_proj",
|
| 1246 |
+
"model.layers.42.mlp.experts.109.up_proj",
|
| 1247 |
+
"model.layers.42.mlp.experts.11.up_proj",
|
| 1248 |
+
"model.layers.42.mlp.experts.110.up_proj",
|
| 1249 |
+
"model.layers.42.mlp.experts.111.up_proj",
|
| 1250 |
+
"model.layers.42.mlp.experts.112.up_proj",
|
| 1251 |
+
"model.layers.42.mlp.experts.113.up_proj",
|
| 1252 |
+
"model.layers.42.mlp.experts.114.up_proj",
|
| 1253 |
+
"model.layers.42.mlp.experts.115.up_proj",
|
| 1254 |
+
"model.layers.42.mlp.experts.116.up_proj",
|
| 1255 |
+
"model.layers.42.mlp.experts.117.up_proj",
|
| 1256 |
+
"model.layers.42.mlp.experts.118.up_proj",
|
| 1257 |
+
"model.layers.42.mlp.experts.119.up_proj",
|
| 1258 |
+
"model.layers.42.mlp.experts.12.up_proj",
|
| 1259 |
+
"model.layers.42.mlp.experts.120.up_proj",
|
| 1260 |
+
"model.layers.42.mlp.experts.121.up_proj",
|
| 1261 |
+
"model.layers.42.mlp.experts.122.up_proj",
|
| 1262 |
+
"model.layers.42.mlp.experts.123.up_proj",
|
| 1263 |
+
"model.layers.42.mlp.experts.124.up_proj",
|
| 1264 |
+
"model.layers.42.mlp.experts.125.up_proj",
|
| 1265 |
+
"model.layers.42.mlp.experts.126.up_proj",
|
| 1266 |
+
"model.layers.42.mlp.experts.127.up_proj",
|
| 1267 |
+
"model.layers.42.mlp.experts.128.up_proj",
|
| 1268 |
+
"model.layers.42.mlp.experts.129.up_proj",
|
| 1269 |
+
"model.layers.42.mlp.experts.13.up_proj",
|
| 1270 |
+
"model.layers.42.mlp.experts.130.up_proj",
|
| 1271 |
+
"model.layers.42.mlp.experts.131.up_proj",
|
| 1272 |
+
"model.layers.42.mlp.experts.132.up_proj",
|
| 1273 |
+
"model.layers.42.mlp.experts.133.up_proj",
|
| 1274 |
+
"model.layers.42.mlp.experts.134.up_proj",
|
| 1275 |
+
"model.layers.42.mlp.experts.135.up_proj",
|
| 1276 |
+
"model.layers.42.mlp.experts.136.up_proj",
|
| 1277 |
+
"model.layers.42.mlp.experts.137.up_proj",
|
| 1278 |
+
"model.layers.42.mlp.experts.138.up_proj",
|
| 1279 |
+
"model.layers.42.mlp.experts.139.up_proj",
|
| 1280 |
+
"model.layers.42.mlp.experts.14.up_proj",
|
| 1281 |
+
"model.layers.42.mlp.experts.140.up_proj",
|
| 1282 |
+
"model.layers.42.mlp.experts.141.up_proj",
|
| 1283 |
+
"model.layers.42.mlp.experts.142.up_proj",
|
| 1284 |
+
"model.layers.42.mlp.experts.143.up_proj",
|
| 1285 |
+
"model.layers.42.mlp.experts.144.up_proj",
|
| 1286 |
+
"model.layers.42.mlp.experts.145.up_proj",
|
| 1287 |
+
"model.layers.42.mlp.experts.146.up_proj",
|
| 1288 |
+
"model.layers.42.mlp.experts.147.up_proj",
|
| 1289 |
+
"model.layers.42.mlp.experts.148.up_proj",
|
| 1290 |
+
"model.layers.42.mlp.experts.149.up_proj",
|
| 1291 |
+
"model.layers.42.mlp.experts.15.up_proj",
|
| 1292 |
+
"model.layers.42.mlp.experts.150.up_proj",
|
| 1293 |
+
"model.layers.42.mlp.experts.151.up_proj",
|
| 1294 |
+
"model.layers.42.mlp.experts.152.up_proj",
|
| 1295 |
+
"model.layers.42.mlp.experts.153.up_proj",
|
| 1296 |
+
"model.layers.42.mlp.experts.154.up_proj",
|
| 1297 |
+
"model.layers.42.mlp.experts.155.up_proj",
|
| 1298 |
+
"model.layers.42.mlp.experts.156.up_proj",
|
| 1299 |
+
"model.layers.42.mlp.experts.157.up_proj",
|
| 1300 |
+
"model.layers.42.mlp.experts.158.up_proj",
|
| 1301 |
+
"model.layers.42.mlp.experts.159.up_proj",
|
| 1302 |
+
"model.layers.42.mlp.experts.16.up_proj",
|
| 1303 |
+
"model.layers.42.mlp.experts.160.up_proj",
|
| 1304 |
+
"model.layers.42.mlp.experts.161.up_proj",
|
| 1305 |
+
"model.layers.42.mlp.experts.162.up_proj",
|
| 1306 |
+
"model.layers.42.mlp.experts.163.up_proj",
|
| 1307 |
+
"model.layers.42.mlp.experts.164.up_proj",
|
| 1308 |
+
"model.layers.42.mlp.experts.165.up_proj",
|
| 1309 |
+
"model.layers.42.mlp.experts.166.up_proj",
|
| 1310 |
+
"model.layers.42.mlp.experts.167.up_proj",
|
| 1311 |
+
"model.layers.42.mlp.experts.168.up_proj",
|
| 1312 |
+
"model.layers.42.mlp.experts.169.up_proj",
|
| 1313 |
+
"model.layers.42.mlp.experts.17.up_proj",
|
| 1314 |
+
"model.layers.42.mlp.experts.170.up_proj",
|
| 1315 |
+
"model.layers.42.mlp.experts.171.up_proj",
|
| 1316 |
+
"model.layers.42.mlp.experts.172.up_proj",
|
| 1317 |
+
"model.layers.42.mlp.experts.173.up_proj",
|
| 1318 |
+
"model.layers.42.mlp.experts.174.up_proj",
|
| 1319 |
+
"model.layers.42.mlp.experts.175.up_proj",
|
| 1320 |
+
"model.layers.42.mlp.experts.176.up_proj",
|
| 1321 |
+
"model.layers.42.mlp.experts.177.up_proj",
|
| 1322 |
+
"model.layers.42.mlp.experts.178.up_proj",
|
| 1323 |
+
"model.layers.42.mlp.experts.179.up_proj",
|
| 1324 |
+
"model.layers.42.mlp.experts.18.up_proj",
|
| 1325 |
+
"model.layers.42.mlp.experts.180.up_proj",
|
| 1326 |
+
"model.layers.42.mlp.experts.181.up_proj",
|
| 1327 |
+
"model.layers.42.mlp.experts.182.up_proj",
|
| 1328 |
+
"model.layers.42.mlp.experts.183.up_proj",
|
| 1329 |
+
"model.layers.42.mlp.experts.184.up_proj",
|
| 1330 |
+
"model.layers.42.mlp.experts.185.up_proj",
|
| 1331 |
+
"model.layers.42.mlp.experts.186.up_proj",
|
| 1332 |
+
"model.layers.42.mlp.experts.187.up_proj",
|
| 1333 |
+
"model.layers.42.mlp.experts.188.up_proj",
|
| 1334 |
+
"model.layers.42.mlp.experts.189.up_proj",
|
| 1335 |
+
"model.layers.42.mlp.experts.19.up_proj",
|
| 1336 |
+
"model.layers.42.mlp.experts.190.up_proj",
|
| 1337 |
+
"model.layers.42.mlp.experts.191.up_proj",
|
| 1338 |
+
"model.layers.42.mlp.experts.192.up_proj",
|
| 1339 |
+
"model.layers.42.mlp.experts.193.up_proj",
|
| 1340 |
+
"model.layers.42.mlp.experts.194.up_proj",
|
| 1341 |
+
"model.layers.42.mlp.experts.195.up_proj",
|
| 1342 |
+
"model.layers.42.mlp.experts.196.up_proj",
|
| 1343 |
+
"model.layers.42.mlp.experts.197.up_proj",
|
| 1344 |
+
"model.layers.42.mlp.experts.198.up_proj",
|
| 1345 |
+
"model.layers.42.mlp.experts.199.up_proj",
|
| 1346 |
+
"model.layers.42.mlp.experts.2.up_proj",
|
| 1347 |
+
"model.layers.42.mlp.experts.20.up_proj",
|
| 1348 |
+
"model.layers.42.mlp.experts.200.up_proj",
|
| 1349 |
+
"model.layers.42.mlp.experts.201.up_proj",
|
| 1350 |
+
"model.layers.42.mlp.experts.202.up_proj",
|
| 1351 |
+
"model.layers.42.mlp.experts.203.up_proj",
|
| 1352 |
+
"model.layers.42.mlp.experts.204.up_proj",
|
| 1353 |
+
"model.layers.42.mlp.experts.205.up_proj",
|
| 1354 |
+
"model.layers.42.mlp.experts.206.up_proj",
|
| 1355 |
+
"model.layers.42.mlp.experts.207.up_proj",
|
| 1356 |
+
"model.layers.42.mlp.experts.208.up_proj",
|
| 1357 |
+
"model.layers.42.mlp.experts.209.up_proj",
|
| 1358 |
+
"model.layers.42.mlp.experts.21.up_proj",
|
| 1359 |
+
"model.layers.42.mlp.experts.210.up_proj",
|
| 1360 |
+
"model.layers.42.mlp.experts.211.up_proj",
|
| 1361 |
+
"model.layers.42.mlp.experts.212.up_proj",
|
| 1362 |
+
"model.layers.42.mlp.experts.213.up_proj",
|
| 1363 |
+
"model.layers.42.mlp.experts.214.up_proj",
|
| 1364 |
+
"model.layers.42.mlp.experts.215.up_proj",
|
| 1365 |
+
"model.layers.42.mlp.experts.216.up_proj",
|
| 1366 |
+
"model.layers.42.mlp.experts.217.up_proj",
|
| 1367 |
+
"model.layers.42.mlp.experts.218.up_proj",
|
| 1368 |
+
"model.layers.42.mlp.experts.219.up_proj",
|
| 1369 |
+
"model.layers.42.mlp.experts.22.up_proj",
|
| 1370 |
+
"model.layers.42.mlp.experts.220.up_proj",
|
| 1371 |
+
"model.layers.42.mlp.experts.221.up_proj",
|
| 1372 |
+
"model.layers.42.mlp.experts.222.up_proj",
|
| 1373 |
+
"model.layers.42.mlp.experts.223.up_proj",
|
| 1374 |
+
"model.layers.42.mlp.experts.224.up_proj",
|
| 1375 |
+
"model.layers.42.mlp.experts.225.up_proj",
|
| 1376 |
+
"model.layers.42.mlp.experts.226.up_proj",
|
| 1377 |
+
"model.layers.42.mlp.experts.227.up_proj",
|
| 1378 |
+
"model.layers.42.mlp.experts.228.up_proj",
|
| 1379 |
+
"model.layers.42.mlp.experts.229.up_proj",
|
| 1380 |
+
"model.layers.42.mlp.experts.23.up_proj",
|
| 1381 |
+
"model.layers.42.mlp.experts.230.up_proj",
|
| 1382 |
+
"model.layers.42.mlp.experts.231.up_proj",
|
| 1383 |
+
"model.layers.42.mlp.experts.232.up_proj",
|
| 1384 |
+
"model.layers.42.mlp.experts.233.up_proj",
|
| 1385 |
+
"model.layers.42.mlp.experts.234.up_proj",
|
| 1386 |
+
"model.layers.42.mlp.experts.235.up_proj",
|
| 1387 |
+
"model.layers.42.mlp.experts.236.up_proj",
|
| 1388 |
+
"model.layers.42.mlp.experts.237.up_proj",
|
| 1389 |
+
"model.layers.42.mlp.experts.238.up_proj",
|
| 1390 |
+
"model.layers.42.mlp.experts.239.up_proj",
|
| 1391 |
+
"model.layers.42.mlp.experts.24.up_proj",
|
| 1392 |
+
"model.layers.42.mlp.experts.240.up_proj",
|
| 1393 |
+
"model.layers.42.mlp.experts.241.up_proj",
|
| 1394 |
+
"model.layers.42.mlp.experts.242.up_proj",
|
| 1395 |
+
"model.layers.42.mlp.experts.243.up_proj",
|
| 1396 |
+
"model.layers.42.mlp.experts.244.up_proj",
|
| 1397 |
+
"model.layers.42.mlp.experts.245.up_proj",
|
| 1398 |
+
"model.layers.42.mlp.experts.246.up_proj",
|
| 1399 |
+
"model.layers.42.mlp.experts.247.up_proj",
|
| 1400 |
+
"model.layers.42.mlp.experts.248.up_proj",
|
| 1401 |
+
"model.layers.42.mlp.experts.249.up_proj",
|
| 1402 |
+
"model.layers.42.mlp.experts.25.up_proj",
|
| 1403 |
+
"model.layers.42.mlp.experts.250.up_proj",
|
| 1404 |
+
"model.layers.42.mlp.experts.251.up_proj",
|
| 1405 |
+
"model.layers.42.mlp.experts.252.up_proj",
|
| 1406 |
+
"model.layers.42.mlp.experts.253.up_proj",
|
| 1407 |
+
"model.layers.42.mlp.experts.254.up_proj",
|
| 1408 |
+
"model.layers.42.mlp.experts.255.up_proj",
|
| 1409 |
+
"model.layers.42.mlp.experts.256.up_proj",
|
| 1410 |
+
"model.layers.42.mlp.experts.257.up_proj",
|
| 1411 |
+
"model.layers.42.mlp.experts.258.up_proj",
|
| 1412 |
+
"model.layers.42.mlp.experts.259.up_proj",
|
| 1413 |
+
"model.layers.42.mlp.experts.26.up_proj",
|
| 1414 |
+
"model.layers.42.mlp.experts.260.up_proj",
|
| 1415 |
+
"model.layers.42.mlp.experts.261.up_proj",
|
| 1416 |
+
"model.layers.42.mlp.experts.262.up_proj",
|
| 1417 |
+
"model.layers.42.mlp.experts.263.up_proj",
|
| 1418 |
+
"model.layers.42.mlp.experts.264.up_proj",
|
| 1419 |
+
"model.layers.42.mlp.experts.265.up_proj",
|
| 1420 |
+
"model.layers.42.mlp.experts.266.up_proj",
|
| 1421 |
+
"model.layers.42.mlp.experts.267.up_proj",
|
| 1422 |
+
"model.layers.42.mlp.experts.268.up_proj",
|
| 1423 |
+
"model.layers.42.mlp.experts.269.up_proj",
|
| 1424 |
+
"model.layers.42.mlp.experts.27.up_proj",
|
| 1425 |
+
"model.layers.42.mlp.experts.270.up_proj",
|
| 1426 |
+
"model.layers.42.mlp.experts.271.up_proj",
|
| 1427 |
+
"model.layers.42.mlp.experts.272.up_proj",
|
| 1428 |
+
"model.layers.42.mlp.experts.273.up_proj",
|
| 1429 |
+
"model.layers.42.mlp.experts.274.up_proj",
|
| 1430 |
+
"model.layers.42.mlp.experts.275.up_proj",
|
| 1431 |
+
"model.layers.42.mlp.experts.276.up_proj",
|
| 1432 |
+
"model.layers.42.mlp.experts.277.up_proj",
|
| 1433 |
+
"model.layers.42.mlp.experts.278.up_proj",
|
| 1434 |
+
"model.layers.42.mlp.experts.279.up_proj",
|
| 1435 |
+
"model.layers.42.mlp.experts.28.up_proj",
|
| 1436 |
+
"model.layers.42.mlp.experts.280.up_proj",
|
| 1437 |
+
"model.layers.42.mlp.experts.281.up_proj",
|
| 1438 |
+
"model.layers.42.mlp.experts.282.up_proj",
|
| 1439 |
+
"model.layers.42.mlp.experts.283.up_proj",
|
| 1440 |
+
"model.layers.42.mlp.experts.284.up_proj",
|
| 1441 |
+
"model.layers.42.mlp.experts.285.up_proj",
|
| 1442 |
+
"model.layers.42.mlp.experts.286.up_proj",
|
| 1443 |
+
"model.layers.42.mlp.experts.287.up_proj",
|
| 1444 |
+
"model.layers.42.mlp.experts.288.up_proj",
|
| 1445 |
+
"model.layers.42.mlp.experts.289.up_proj",
|
| 1446 |
+
"model.layers.42.mlp.experts.29.up_proj",
|
| 1447 |
+
"model.layers.42.mlp.experts.290.up_proj",
|
| 1448 |
+
"model.layers.42.mlp.experts.291.up_proj",
|
| 1449 |
+
"model.layers.42.mlp.experts.292.up_proj",
|
| 1450 |
+
"model.layers.42.mlp.experts.293.up_proj",
|
| 1451 |
+
"model.layers.42.mlp.experts.294.up_proj",
|
| 1452 |
+
"model.layers.42.mlp.experts.295.up_proj",
|
| 1453 |
+
"model.layers.42.mlp.experts.296.up_proj",
|
| 1454 |
+
"model.layers.42.mlp.experts.297.up_proj",
|
| 1455 |
+
"model.layers.42.mlp.experts.298.up_proj",
|
| 1456 |
+
"model.layers.42.mlp.experts.299.up_proj",
|
| 1457 |
+
"model.layers.42.mlp.experts.3.up_proj",
|
| 1458 |
+
"model.layers.42.mlp.experts.30.up_proj",
|
| 1459 |
+
"model.layers.42.mlp.experts.300.up_proj",
|
| 1460 |
+
"model.layers.42.mlp.experts.301.up_proj",
|
| 1461 |
+
"model.layers.42.mlp.experts.302.up_proj",
|
| 1462 |
+
"model.layers.42.mlp.experts.303.up_proj",
|
| 1463 |
+
"model.layers.42.mlp.experts.304.up_proj",
|
| 1464 |
+
"model.layers.42.mlp.experts.305.up_proj",
|
| 1465 |
+
"model.layers.42.mlp.experts.306.up_proj",
|
| 1466 |
+
"model.layers.42.mlp.experts.307.up_proj",
|
| 1467 |
+
"model.layers.42.mlp.experts.308.up_proj",
|
| 1468 |
+
"model.layers.42.mlp.experts.309.up_proj",
|
| 1469 |
+
"model.layers.42.mlp.experts.31.up_proj",
|
| 1470 |
+
"model.layers.42.mlp.experts.310.up_proj",
|
| 1471 |
+
"model.layers.42.mlp.experts.311.up_proj",
|
| 1472 |
+
"model.layers.42.mlp.experts.312.up_proj",
|
| 1473 |
+
"model.layers.42.mlp.experts.313.up_proj",
|
| 1474 |
+
"model.layers.42.mlp.experts.314.up_proj",
|
| 1475 |
+
"model.layers.42.mlp.experts.315.up_proj",
|
| 1476 |
+
"model.layers.42.mlp.experts.316.up_proj",
|
| 1477 |
+
"model.layers.42.mlp.experts.317.up_proj",
|
| 1478 |
+
"model.layers.42.mlp.experts.318.up_proj",
|
| 1479 |
+
"model.layers.42.mlp.experts.319.up_proj",
|
| 1480 |
+
"model.layers.42.mlp.experts.32.up_proj",
|
| 1481 |
+
"model.layers.42.mlp.experts.320.up_proj",
|
| 1482 |
+
"model.layers.42.mlp.experts.321.up_proj",
|
| 1483 |
+
"model.layers.42.mlp.experts.322.up_proj",
|
| 1484 |
+
"model.layers.42.mlp.experts.323.up_proj",
|
| 1485 |
+
"model.layers.42.mlp.experts.324.up_proj",
|
| 1486 |
+
"model.layers.42.mlp.experts.325.up_proj",
|
| 1487 |
+
"model.layers.42.mlp.experts.326.up_proj",
|
| 1488 |
+
"model.layers.42.mlp.experts.327.up_proj",
|
| 1489 |
+
"model.layers.42.mlp.experts.328.up_proj",
|
| 1490 |
+
"model.layers.42.mlp.experts.329.up_proj",
|
| 1491 |
+
"model.layers.42.mlp.experts.33.up_proj",
|
| 1492 |
+
"model.layers.42.mlp.experts.330.up_proj",
|
| 1493 |
+
"model.layers.42.mlp.experts.331.up_proj",
|
| 1494 |
+
"model.layers.42.mlp.experts.332.up_proj",
|
| 1495 |
+
"model.layers.42.mlp.experts.333.up_proj",
|
| 1496 |
+
"model.layers.42.mlp.experts.334.up_proj",
|
| 1497 |
+
"model.layers.42.mlp.experts.335.up_proj",
|
| 1498 |
+
"model.layers.42.mlp.experts.336.up_proj",
|
| 1499 |
+
"model.layers.42.mlp.experts.337.up_proj",
|
| 1500 |
+
"model.layers.42.mlp.experts.338.up_proj",
|
| 1501 |
+
"model.layers.42.mlp.experts.339.up_proj",
|
| 1502 |
+
"model.layers.42.mlp.experts.34.up_proj",
|
| 1503 |
+
"model.layers.42.mlp.experts.340.up_proj",
|
| 1504 |
+
"model.layers.42.mlp.experts.341.up_proj",
|
| 1505 |
+
"model.layers.42.mlp.experts.342.up_proj",
|
| 1506 |
+
"model.layers.42.mlp.experts.343.up_proj",
|
| 1507 |
+
"model.layers.42.mlp.experts.344.up_proj",
|
| 1508 |
+
"model.layers.42.mlp.experts.345.up_proj",
|
| 1509 |
+
"model.layers.42.mlp.experts.346.up_proj",
|
| 1510 |
+
"model.layers.42.mlp.experts.347.up_proj",
|
| 1511 |
+
"model.layers.42.mlp.experts.348.up_proj",
|
| 1512 |
+
"model.layers.42.mlp.experts.349.up_proj",
|
| 1513 |
+
"model.layers.42.mlp.experts.35.up_proj",
|
| 1514 |
+
"model.layers.42.mlp.experts.350.up_proj",
|
| 1515 |
+
"model.layers.42.mlp.experts.351.up_proj",
|
| 1516 |
+
"model.layers.42.mlp.experts.352.up_proj",
|
| 1517 |
+
"model.layers.42.mlp.experts.353.up_proj",
|
| 1518 |
+
"model.layers.42.mlp.experts.354.up_proj",
|
| 1519 |
+
"model.layers.42.mlp.experts.355.up_proj",
|
| 1520 |
+
"model.layers.42.mlp.experts.356.up_proj",
|
| 1521 |
+
"model.layers.42.mlp.experts.357.up_proj",
|
| 1522 |
+
"model.layers.42.mlp.experts.358.up_proj",
|
| 1523 |
+
"model.layers.42.mlp.experts.359.up_proj",
|
| 1524 |
+
"model.layers.42.mlp.experts.36.up_proj",
|
| 1525 |
+
"model.layers.42.mlp.experts.360.up_proj",
|
| 1526 |
+
"model.layers.42.mlp.experts.361.up_proj",
|
| 1527 |
+
"model.layers.42.mlp.experts.362.up_proj",
|
| 1528 |
+
"model.layers.42.mlp.experts.363.up_proj",
|
| 1529 |
+
"model.layers.42.mlp.experts.364.up_proj",
|
| 1530 |
+
"model.layers.42.mlp.experts.365.up_proj",
|
| 1531 |
+
"model.layers.42.mlp.experts.366.up_proj",
|
| 1532 |
+
"model.layers.42.mlp.experts.367.up_proj",
|
| 1533 |
+
"model.layers.42.mlp.experts.368.up_proj",
|
| 1534 |
+
"model.layers.42.mlp.experts.369.up_proj",
|
| 1535 |
+
"model.layers.42.mlp.experts.37.up_proj",
|
| 1536 |
+
"model.layers.42.mlp.experts.370.up_proj",
|
| 1537 |
+
"model.layers.42.mlp.experts.371.up_proj",
|
| 1538 |
+
"model.layers.42.mlp.experts.372.up_proj",
|
| 1539 |
+
"model.layers.42.mlp.experts.373.up_proj",
|
| 1540 |
+
"model.layers.42.mlp.experts.374.up_proj",
|
| 1541 |
+
"model.layers.42.mlp.experts.375.up_proj",
|
| 1542 |
+
"model.layers.42.mlp.experts.376.up_proj",
|
| 1543 |
+
"model.layers.42.mlp.experts.377.up_proj",
|
| 1544 |
+
"model.layers.42.mlp.experts.378.up_proj",
|
| 1545 |
+
"model.layers.42.mlp.experts.379.up_proj",
|
| 1546 |
+
"model.layers.42.mlp.experts.38.up_proj",
|
| 1547 |
+
"model.layers.42.mlp.experts.380.up_proj",
|
| 1548 |
+
"model.layers.42.mlp.experts.381.up_proj",
|
| 1549 |
+
"model.layers.42.mlp.experts.382.up_proj",
|
| 1550 |
+
"model.layers.42.mlp.experts.383.up_proj",
|
| 1551 |
+
"model.layers.42.mlp.experts.384.up_proj",
|
| 1552 |
+
"model.layers.42.mlp.experts.385.up_proj",
|
| 1553 |
+
"model.layers.42.mlp.experts.386.up_proj",
|
| 1554 |
+
"model.layers.42.mlp.experts.387.up_proj",
|
| 1555 |
+
"model.layers.42.mlp.experts.388.up_proj",
|
| 1556 |
+
"model.layers.42.mlp.experts.389.up_proj",
|
| 1557 |
+
"model.layers.42.mlp.experts.39.up_proj",
|
| 1558 |
+
"model.layers.42.mlp.experts.390.up_proj",
|
| 1559 |
+
"model.layers.42.mlp.experts.391.up_proj",
|
| 1560 |
+
"model.layers.42.mlp.experts.392.up_proj",
|
| 1561 |
+
"model.layers.42.mlp.experts.393.up_proj",
|
| 1562 |
+
"model.layers.42.mlp.experts.394.up_proj",
|
| 1563 |
+
"model.layers.42.mlp.experts.395.up_proj",
|
| 1564 |
+
"model.layers.42.mlp.experts.396.up_proj",
|
| 1565 |
+
"model.layers.42.mlp.experts.397.up_proj",
|
| 1566 |
+
"model.layers.42.mlp.experts.398.up_proj",
|
| 1567 |
+
"model.layers.42.mlp.experts.399.up_proj",
|
| 1568 |
+
"model.layers.42.mlp.experts.4.up_proj",
|
| 1569 |
+
"model.layers.42.mlp.experts.40.up_proj",
|
| 1570 |
+
"model.layers.42.mlp.experts.400.up_proj",
|
| 1571 |
+
"model.layers.42.mlp.experts.401.up_proj",
|
| 1572 |
+
"model.layers.42.mlp.experts.402.up_proj",
|
| 1573 |
+
"model.layers.42.mlp.experts.403.up_proj",
|
| 1574 |
+
"model.layers.42.mlp.experts.404.up_proj",
|
| 1575 |
+
"model.layers.42.mlp.experts.405.up_proj",
|
| 1576 |
+
"model.layers.42.mlp.experts.406.up_proj",
|
| 1577 |
+
"model.layers.42.mlp.experts.407.up_proj",
|
| 1578 |
+
"model.layers.42.mlp.experts.408.up_proj",
|
| 1579 |
+
"model.layers.42.mlp.experts.409.up_proj",
|
| 1580 |
+
"model.layers.42.mlp.experts.41.up_proj",
|
| 1581 |
+
"model.layers.42.mlp.experts.410.up_proj",
|
| 1582 |
+
"model.layers.42.mlp.experts.411.up_proj",
|
| 1583 |
+
"model.layers.42.mlp.experts.412.up_proj",
|
| 1584 |
+
"model.layers.42.mlp.experts.413.up_proj",
|
| 1585 |
+
"model.layers.42.mlp.experts.414.up_proj",
|
| 1586 |
+
"model.layers.42.mlp.experts.415.up_proj",
|
| 1587 |
+
"model.layers.42.mlp.experts.416.up_proj",
|
| 1588 |
+
"model.layers.42.mlp.experts.417.up_proj",
|
| 1589 |
+
"model.layers.42.mlp.experts.418.up_proj",
|
| 1590 |
+
"model.layers.42.mlp.experts.419.up_proj",
|
| 1591 |
+
"model.layers.42.mlp.experts.42.up_proj",
|
| 1592 |
+
"model.layers.42.mlp.experts.420.up_proj",
|
| 1593 |
+
"model.layers.42.mlp.experts.421.up_proj",
|
| 1594 |
+
"model.layers.42.mlp.experts.422.up_proj",
|
| 1595 |
+
"model.layers.42.mlp.experts.423.up_proj",
|
| 1596 |
+
"model.layers.42.mlp.experts.424.up_proj",
|
| 1597 |
+
"model.layers.42.mlp.experts.425.up_proj",
|
| 1598 |
+
"model.layers.42.mlp.experts.426.up_proj",
|
| 1599 |
+
"model.layers.42.mlp.experts.427.up_proj",
|
| 1600 |
+
"model.layers.42.mlp.experts.428.up_proj",
|
| 1601 |
+
"model.layers.42.mlp.experts.429.up_proj",
|
| 1602 |
+
"model.layers.42.mlp.experts.43.up_proj",
|
| 1603 |
+
"model.layers.42.mlp.experts.430.up_proj",
|
| 1604 |
+
"model.layers.42.mlp.experts.431.up_proj",
|
| 1605 |
+
"model.layers.42.mlp.experts.432.up_proj",
|
| 1606 |
+
"model.layers.42.mlp.experts.433.up_proj",
|
| 1607 |
+
"model.layers.42.mlp.experts.434.up_proj",
|
| 1608 |
+
"model.layers.42.mlp.experts.435.up_proj",
|
| 1609 |
+
"model.layers.42.mlp.experts.436.up_proj",
|
| 1610 |
+
"model.layers.42.mlp.experts.437.up_proj",
|
| 1611 |
+
"model.layers.42.mlp.experts.438.up_proj",
|
| 1612 |
+
"model.layers.42.mlp.experts.439.up_proj",
|
| 1613 |
+
"model.layers.42.mlp.experts.44.up_proj",
|
| 1614 |
+
"model.layers.42.mlp.experts.440.up_proj",
|
| 1615 |
+
"model.layers.42.mlp.experts.441.up_proj",
|
| 1616 |
+
"model.layers.42.mlp.experts.442.up_proj",
|
| 1617 |
+
"model.layers.42.mlp.experts.443.up_proj",
|
| 1618 |
+
"model.layers.42.mlp.experts.444.up_proj",
|
| 1619 |
+
"model.layers.42.mlp.experts.445.up_proj",
|
| 1620 |
+
"model.layers.42.mlp.experts.446.up_proj",
|
| 1621 |
+
"model.layers.42.mlp.experts.447.up_proj",
|
| 1622 |
+
"model.layers.42.mlp.experts.448.up_proj",
|
| 1623 |
+
"model.layers.42.mlp.experts.449.up_proj",
|
| 1624 |
+
"model.layers.42.mlp.experts.45.up_proj",
|
| 1625 |
+
"model.layers.42.mlp.experts.450.up_proj",
|
| 1626 |
+
"model.layers.42.mlp.experts.451.up_proj",
|
| 1627 |
+
"model.layers.42.mlp.experts.452.up_proj",
|
| 1628 |
+
"model.layers.42.mlp.experts.453.up_proj",
|
| 1629 |
+
"model.layers.42.mlp.experts.454.up_proj",
|
| 1630 |
+
"model.layers.42.mlp.experts.455.up_proj",
|
| 1631 |
+
"model.layers.42.mlp.experts.456.up_proj",
|
| 1632 |
+
"model.layers.42.mlp.experts.457.up_proj",
|
| 1633 |
+
"model.layers.42.mlp.experts.458.up_proj",
|
| 1634 |
+
"model.layers.42.mlp.experts.459.up_proj",
|
| 1635 |
+
"model.layers.42.mlp.experts.46.up_proj",
|
| 1636 |
+
"model.layers.42.mlp.experts.460.up_proj",
|
| 1637 |
+
"model.layers.42.mlp.experts.461.up_proj",
|
| 1638 |
+
"model.layers.42.mlp.experts.462.up_proj",
|
| 1639 |
+
"model.layers.42.mlp.experts.463.up_proj",
|
| 1640 |
+
"model.layers.42.mlp.experts.464.up_proj",
|
| 1641 |
+
"model.layers.42.mlp.experts.465.up_proj",
|
| 1642 |
+
"model.layers.42.mlp.experts.466.up_proj",
|
| 1643 |
+
"model.layers.42.mlp.experts.467.up_proj",
|
| 1644 |
+
"model.layers.42.mlp.experts.468.up_proj",
|
| 1645 |
+
"model.layers.42.mlp.experts.469.up_proj",
|
| 1646 |
+
"model.layers.42.mlp.experts.47.up_proj",
|
| 1647 |
+
"model.layers.42.mlp.experts.470.up_proj",
|
| 1648 |
+
"model.layers.42.mlp.experts.471.up_proj",
|
| 1649 |
+
"model.layers.42.mlp.experts.472.up_proj",
|
| 1650 |
+
"model.layers.42.mlp.experts.473.up_proj",
|
| 1651 |
+
"model.layers.42.mlp.experts.474.up_proj",
|
| 1652 |
+
"model.layers.42.mlp.experts.475.up_proj",
|
| 1653 |
+
"model.layers.42.mlp.experts.476.up_proj",
|
| 1654 |
+
"model.layers.42.mlp.experts.477.up_proj",
|
| 1655 |
+
"model.layers.42.mlp.experts.478.up_proj",
|
| 1656 |
+
"model.layers.42.mlp.experts.479.up_proj",
|
| 1657 |
+
"model.layers.42.mlp.experts.48.up_proj",
|
| 1658 |
+
"model.layers.42.mlp.experts.480.up_proj",
|
| 1659 |
+
"model.layers.42.mlp.experts.481.up_proj",
|
| 1660 |
+
"model.layers.42.mlp.experts.482.up_proj",
|
| 1661 |
+
"model.layers.42.mlp.experts.483.up_proj",
|
| 1662 |
+
"model.layers.42.mlp.experts.484.up_proj",
|
| 1663 |
+
"model.layers.42.mlp.experts.485.up_proj",
|
| 1664 |
+
"model.layers.42.mlp.experts.486.up_proj",
|
| 1665 |
+
"model.layers.42.mlp.experts.487.up_proj",
|
| 1666 |
+
"model.layers.42.mlp.experts.488.up_proj",
|
| 1667 |
+
"model.layers.42.mlp.experts.489.up_proj",
|
| 1668 |
+
"model.layers.42.mlp.experts.49.up_proj",
|
| 1669 |
+
"model.layers.42.mlp.experts.490.up_proj",
|
| 1670 |
+
"model.layers.42.mlp.experts.491.up_proj",
|
| 1671 |
+
"model.layers.42.mlp.experts.492.up_proj",
|
| 1672 |
+
"model.layers.42.mlp.experts.493.up_proj",
|
| 1673 |
+
"model.layers.42.mlp.experts.494.up_proj",
|
| 1674 |
+
"model.layers.42.mlp.experts.495.up_proj",
|
| 1675 |
+
"model.layers.42.mlp.experts.496.up_proj",
|
| 1676 |
+
"model.layers.42.mlp.experts.497.up_proj",
|
| 1677 |
+
"model.layers.42.mlp.experts.498.up_proj",
|
| 1678 |
+
"model.layers.42.mlp.experts.499.up_proj",
|
| 1679 |
+
"model.layers.42.mlp.experts.5.up_proj",
|
| 1680 |
+
"model.layers.42.mlp.experts.50.up_proj",
|
| 1681 |
+
"model.layers.42.mlp.experts.500.up_proj",
|
| 1682 |
+
"model.layers.42.mlp.experts.501.up_proj",
|
| 1683 |
+
"model.layers.42.mlp.experts.502.up_proj",
|
| 1684 |
+
"model.layers.42.mlp.experts.503.up_proj",
|
| 1685 |
+
"model.layers.42.mlp.experts.504.up_proj",
|
| 1686 |
+
"model.layers.42.mlp.experts.505.up_proj",
|
| 1687 |
+
"model.layers.42.mlp.experts.506.up_proj",
|
| 1688 |
+
"model.layers.42.mlp.experts.507.up_proj",
|
| 1689 |
+
"model.layers.42.mlp.experts.508.up_proj",
|
| 1690 |
+
"model.layers.42.mlp.experts.509.up_proj",
|
| 1691 |
+
"model.layers.42.mlp.experts.51.up_proj",
|
| 1692 |
+
"model.layers.42.mlp.experts.510.up_proj",
|
| 1693 |
+
"model.layers.42.mlp.experts.511.up_proj",
|
| 1694 |
+
"model.layers.42.mlp.experts.52.up_proj",
|
| 1695 |
+
"model.layers.42.mlp.experts.53.up_proj",
|
| 1696 |
+
"model.layers.42.mlp.experts.54.up_proj",
|
| 1697 |
+
"model.layers.42.mlp.experts.55.up_proj",
|
| 1698 |
+
"model.layers.42.mlp.experts.56.up_proj",
|
| 1699 |
+
"model.layers.42.mlp.experts.57.up_proj",
|
| 1700 |
+
"model.layers.42.mlp.experts.58.up_proj",
|
| 1701 |
+
"model.layers.42.mlp.experts.59.up_proj",
|
| 1702 |
+
"model.layers.42.mlp.experts.6.up_proj",
|
| 1703 |
+
"model.layers.42.mlp.experts.60.up_proj",
|
| 1704 |
+
"model.layers.42.mlp.experts.61.up_proj",
|
| 1705 |
+
"model.layers.42.mlp.experts.62.up_proj",
|
| 1706 |
+
"model.layers.42.mlp.experts.63.up_proj",
|
| 1707 |
+
"model.layers.42.mlp.experts.64.up_proj",
|
| 1708 |
+
"model.layers.42.mlp.experts.65.up_proj",
|
| 1709 |
+
"model.layers.42.mlp.experts.66.up_proj",
|
| 1710 |
+
"model.layers.42.mlp.experts.67.up_proj",
|
| 1711 |
+
"model.layers.42.mlp.experts.68.up_proj",
|
| 1712 |
+
"model.layers.42.mlp.experts.69.up_proj",
|
| 1713 |
+
"model.layers.42.mlp.experts.7.up_proj",
|
| 1714 |
+
"model.layers.42.mlp.experts.70.up_proj",
|
| 1715 |
+
"model.layers.42.mlp.experts.71.up_proj",
|
| 1716 |
+
"model.layers.42.mlp.experts.72.up_proj",
|
| 1717 |
+
"model.layers.42.mlp.experts.73.up_proj",
|
| 1718 |
+
"model.layers.42.mlp.experts.74.up_proj",
|
| 1719 |
+
"model.layers.42.mlp.experts.75.up_proj",
|
| 1720 |
+
"model.layers.42.mlp.experts.76.up_proj",
|
| 1721 |
+
"model.layers.42.mlp.experts.77.up_proj",
|
| 1722 |
+
"model.layers.42.mlp.experts.78.up_proj",
|
| 1723 |
+
"model.layers.42.mlp.experts.79.up_proj",
|
| 1724 |
+
"model.layers.42.mlp.experts.8.up_proj",
|
| 1725 |
+
"model.layers.42.mlp.experts.80.up_proj",
|
| 1726 |
+
"model.layers.42.mlp.experts.81.up_proj",
|
| 1727 |
+
"model.layers.42.mlp.experts.82.up_proj",
|
| 1728 |
+
"model.layers.42.mlp.experts.83.up_proj",
|
| 1729 |
+
"model.layers.42.mlp.experts.84.up_proj",
|
| 1730 |
+
"model.layers.42.mlp.experts.85.up_proj",
|
| 1731 |
+
"model.layers.42.mlp.experts.86.up_proj",
|
| 1732 |
+
"model.layers.42.mlp.experts.87.up_proj",
|
| 1733 |
+
"model.layers.42.mlp.experts.88.up_proj",
|
| 1734 |
+
"model.layers.42.mlp.experts.89.up_proj",
|
| 1735 |
+
"model.layers.42.mlp.experts.9.up_proj",
|
| 1736 |
+
"model.layers.42.mlp.experts.90.up_proj",
|
| 1737 |
+
"model.layers.42.mlp.experts.91.up_proj",
|
| 1738 |
+
"model.layers.42.mlp.experts.92.up_proj",
|
| 1739 |
+
"model.layers.42.mlp.experts.93.up_proj",
|
| 1740 |
+
"model.layers.42.mlp.experts.94.up_proj",
|
| 1741 |
+
"model.layers.42.mlp.experts.95.up_proj",
|
| 1742 |
+
"model.layers.42.mlp.experts.96.up_proj",
|
| 1743 |
+
"model.layers.42.mlp.experts.97.up_proj",
|
| 1744 |
+
"model.layers.42.mlp.experts.98.up_proj",
|
| 1745 |
+
"model.layers.42.mlp.experts.99.up_proj",
|
| 1746 |
"re:.*model\\.word_embeddings$"
|
| 1747 |
]
|
| 1748 |
}
|
vllm_patch/README.md
CHANGED
|
@@ -61,3 +61,24 @@ apply_moe_activation(SILU, x=[10,-10,2,1], clamp_limit=4) -> [ 7.844, -0.00045]
|
|
| 61 |
```
|
| 62 |
|
| 63 |
All four edits are pure Python, so an editable install picks them up with no rebuild.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 61 |
```
|
| 62 |
|
| 63 |
All four edits are pure Python, so an editable install picks them up with no rebuild.
|
| 64 |
+
|
| 65 |
+
---
|
| 66 |
+
|
| 67 |
+
## Additional patch: `unquantized_moe_oracle.py` (NVFP4 + MTP only)
|
| 68 |
+
|
| 69 |
+
Path: `vllm/model_executor/layers/fused_moe/oracle/unquantized.py`
|
| 70 |
+
|
| 71 |
+
`moe_backend` is a single global setting, but an NVFP4 checkpoint with a BF16 MTP layer
|
| 72 |
+
needs two different backends: `marlin` for the NVFP4 routed experts (the only backend
|
| 73 |
+
supporting weight-only NVFP4A16) and something else for the unquantized MTP layer, since
|
| 74 |
+
marlin has no unquantized kernel. Without this patch vLLM raises
|
| 75 |
+
`moe_backend='marlin' is not supported for unquantized MoE`.
|
| 76 |
+
|
| 77 |
+
The patch treats marlin as weight-only — exactly as the code already treats `humming` —
|
| 78 |
+
and routes the unquantized layer to Triton. Triton is chosen over `auto` deliberately:
|
| 79 |
+
auto selects FlashInfer TRT-LLM, whose grouped-routing kernel is warp-limited to 32
|
| 80 |
+
experts per group, and Ling-3.0 has 64 (512 experts / 8 groups). That path either raises
|
| 81 |
+
`Routing kernel expects #experts per group <= warp size (32)` or silently mis-routes,
|
| 82 |
+
producing a drafter the target never agrees with (0% acceptance, no error).
|
| 83 |
+
|
| 84 |
+
Not needed for the MXFP4 build, which can use `moe_backend: triton` globally.
|
vllm_patch/unquantized_moe_oracle.py
ADDED
|
@@ -0,0 +1,501 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# SPDX-License-Identifier: Apache-2.0
|
| 2 |
+
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
| 3 |
+
|
| 4 |
+
from enum import Enum
|
| 5 |
+
from typing import TYPE_CHECKING
|
| 6 |
+
|
| 7 |
+
import torch
|
| 8 |
+
|
| 9 |
+
import vllm.envs as envs
|
| 10 |
+
import vllm.model_executor.layers.fused_moe.modular_kernel as mk
|
| 11 |
+
from vllm._aiter_ops import rocm_aiter_ops
|
| 12 |
+
from vllm.config.kernel import MoEBackend
|
| 13 |
+
from vllm.logger import init_logger
|
| 14 |
+
from vllm.model_executor.layers.fused_moe.activation import MoEActivation
|
| 15 |
+
from vllm.model_executor.layers.fused_moe.all2all_utils import (
|
| 16 |
+
maybe_make_prepare_finalize,
|
| 17 |
+
)
|
| 18 |
+
from vllm.model_executor.layers.fused_moe.config import (
|
| 19 |
+
FusedMoEConfig,
|
| 20 |
+
FusedMoEQuantConfig,
|
| 21 |
+
)
|
| 22 |
+
from vllm.model_executor.layers.fused_moe.oracle.base import MoEKernelOracle
|
| 23 |
+
from vllm.model_executor.layers.quantization.utils.flashinfer_utils import (
|
| 24 |
+
align_moe_weights_for_fi,
|
| 25 |
+
convert_moe_weights_to_flashinfer_trtllm_block_layout,
|
| 26 |
+
swap_w13_to_w31,
|
| 27 |
+
)
|
| 28 |
+
from vllm.platforms import current_platform
|
| 29 |
+
|
| 30 |
+
if TYPE_CHECKING:
|
| 31 |
+
from vllm.model_executor.layers.quantization.utils.quant_utils import QuantKey
|
| 32 |
+
|
| 33 |
+
logger = init_logger(__name__)
|
| 34 |
+
|
| 35 |
+
|
| 36 |
+
class UnquantizedMoeBackend(Enum):
|
| 37 |
+
FLASHINFER_TRTLLM = "FlashInfer TRTLLM"
|
| 38 |
+
FLASHINFER_CUTLASS = "FlashInfer CUTLASS"
|
| 39 |
+
AITER = "ROCm AITER"
|
| 40 |
+
TRITON = "TRITON"
|
| 41 |
+
BATCHED_TRITON = "BATCHED_TRITON"
|
| 42 |
+
CPU = "CPU"
|
| 43 |
+
XPU = "XPU"
|
| 44 |
+
TPU = "TPU"
|
| 45 |
+
OOT = "OOT"
|
| 46 |
+
|
| 47 |
+
|
| 48 |
+
def _get_priority_backends(moe_config: FusedMoEConfig) -> list[UnquantizedMoeBackend]:
|
| 49 |
+
"""
|
| 50 |
+
Get available backends in priority order based on platform and config.
|
| 51 |
+
|
| 52 |
+
This function can be extended to become more complex as needed.
|
| 53 |
+
"""
|
| 54 |
+
|
| 55 |
+
def _move_to_back(
|
| 56 |
+
backends: list[UnquantizedMoeBackend],
|
| 57 |
+
backend: UnquantizedMoeBackend,
|
| 58 |
+
) -> None:
|
| 59 |
+
backends.append(backends.pop(backends.index(backend)))
|
| 60 |
+
|
| 61 |
+
if current_platform.is_rocm():
|
| 62 |
+
_AVAILABLE_BACKENDS = [
|
| 63 |
+
UnquantizedMoeBackend.AITER,
|
| 64 |
+
UnquantizedMoeBackend.TRITON,
|
| 65 |
+
UnquantizedMoeBackend.BATCHED_TRITON,
|
| 66 |
+
]
|
| 67 |
+
elif current_platform.is_cuda():
|
| 68 |
+
_AVAILABLE_BACKENDS = [
|
| 69 |
+
UnquantizedMoeBackend.FLASHINFER_TRTLLM,
|
| 70 |
+
UnquantizedMoeBackend.FLASHINFER_CUTLASS,
|
| 71 |
+
UnquantizedMoeBackend.TRITON,
|
| 72 |
+
UnquantizedMoeBackend.BATCHED_TRITON,
|
| 73 |
+
]
|
| 74 |
+
|
| 75 |
+
# On Hopper (SM90), the FlashInfer unquantized MoE kernels are slower
|
| 76 |
+
# than Triton, so prefer Triton by default.
|
| 77 |
+
if current_platform.is_device_capability_family(90):
|
| 78 |
+
_move_to_back(_AVAILABLE_BACKENDS, UnquantizedMoeBackend.FLASHINFER_TRTLLM)
|
| 79 |
+
_move_to_back(_AVAILABLE_BACKENDS, UnquantizedMoeBackend.FLASHINFER_CUTLASS)
|
| 80 |
+
|
| 81 |
+
# HACK: Qwen3.5 has crash with FLASHINFER_CUTLASS BF16 if DEP.
|
| 82 |
+
# Updating the oracle querying logic is out of the scope of this
|
| 83 |
+
# PR. Need to fix the kernel or update structure in follow up.
|
| 84 |
+
if moe_config.moe_parallel_config.dp_size > 1:
|
| 85 |
+
_move_to_back(_AVAILABLE_BACKENDS, UnquantizedMoeBackend.FLASHINFER_CUTLASS)
|
| 86 |
+
|
| 87 |
+
# HACK: unquantized FlashInfer aliases SWIGLUOAI to plain Swiglu
|
| 88 |
+
# (swiglu_alpha/limit only set on the MXFP4 branch). Route to
|
| 89 |
+
# Triton's swigluoai_and_mul until that's plumbed through. Same
|
| 90 |
+
# demotion pattern as the Qwen3.5/dp_size hack above.
|
| 91 |
+
if moe_config.activation == MoEActivation.SWIGLUOAI:
|
| 92 |
+
_move_to_back(_AVAILABLE_BACKENDS, UnquantizedMoeBackend.FLASHINFER_TRTLLM)
|
| 93 |
+
_move_to_back(_AVAILABLE_BACKENDS, UnquantizedMoeBackend.FLASHINFER_CUTLASS)
|
| 94 |
+
|
| 95 |
+
elif current_platform.is_xpu():
|
| 96 |
+
_AVAILABLE_BACKENDS = [UnquantizedMoeBackend.XPU]
|
| 97 |
+
elif current_platform.is_cpu():
|
| 98 |
+
_AVAILABLE_BACKENDS = [UnquantizedMoeBackend.CPU]
|
| 99 |
+
return _AVAILABLE_BACKENDS
|
| 100 |
+
|
| 101 |
+
|
| 102 |
+
def backend_to_kernel_cls(
|
| 103 |
+
backend: UnquantizedMoeBackend,
|
| 104 |
+
) -> list[type[mk.FusedMoEExperts]]:
|
| 105 |
+
if backend == UnquantizedMoeBackend.FLASHINFER_TRTLLM:
|
| 106 |
+
from vllm.model_executor.layers.fused_moe.experts.trtllm_bf16_moe import (
|
| 107 |
+
TrtLlmBf16ExpertsModular,
|
| 108 |
+
TrtLlmBf16ExpertsMonolithic,
|
| 109 |
+
)
|
| 110 |
+
|
| 111 |
+
return [TrtLlmBf16ExpertsMonolithic, TrtLlmBf16ExpertsModular]
|
| 112 |
+
|
| 113 |
+
elif backend == UnquantizedMoeBackend.FLASHINFER_CUTLASS:
|
| 114 |
+
from vllm.model_executor.layers.fused_moe.experts.flashinfer_cutlass_moe import ( # noqa: E501
|
| 115 |
+
FlashInferExperts,
|
| 116 |
+
)
|
| 117 |
+
|
| 118 |
+
return [FlashInferExperts]
|
| 119 |
+
|
| 120 |
+
elif backend == UnquantizedMoeBackend.AITER:
|
| 121 |
+
from vllm.model_executor.layers.fused_moe.experts.rocm_aiter_moe import (
|
| 122 |
+
AiterExperts,
|
| 123 |
+
)
|
| 124 |
+
|
| 125 |
+
return [AiterExperts]
|
| 126 |
+
|
| 127 |
+
elif backend == UnquantizedMoeBackend.TRITON:
|
| 128 |
+
from vllm.model_executor.layers.fused_moe.experts.triton_moe import (
|
| 129 |
+
TritonExperts,
|
| 130 |
+
)
|
| 131 |
+
|
| 132 |
+
return [TritonExperts]
|
| 133 |
+
|
| 134 |
+
elif backend == UnquantizedMoeBackend.BATCHED_TRITON:
|
| 135 |
+
from vllm.model_executor.layers.fused_moe.experts.fused_batched_moe import (
|
| 136 |
+
BatchedTritonExperts,
|
| 137 |
+
)
|
| 138 |
+
|
| 139 |
+
return [BatchedTritonExperts]
|
| 140 |
+
|
| 141 |
+
elif backend == UnquantizedMoeBackend.XPU:
|
| 142 |
+
from vllm.model_executor.layers.fused_moe.experts.xpu_moe import XPUExperts
|
| 143 |
+
|
| 144 |
+
return [XPUExperts]
|
| 145 |
+
|
| 146 |
+
elif backend == UnquantizedMoeBackend.CPU:
|
| 147 |
+
from vllm.model_executor.layers.fused_moe.experts.cpu_moe import (
|
| 148 |
+
ArmCPUUnquantizedExperts,
|
| 149 |
+
CPUUnquantizedExperts,
|
| 150 |
+
X86CPUUnquantizedExperts,
|
| 151 |
+
)
|
| 152 |
+
|
| 153 |
+
# Prefer architecture-specific kernels before the portable vector path.
|
| 154 |
+
return [
|
| 155 |
+
X86CPUUnquantizedExperts,
|
| 156 |
+
ArmCPUUnquantizedExperts,
|
| 157 |
+
CPUUnquantizedExperts,
|
| 158 |
+
]
|
| 159 |
+
|
| 160 |
+
else:
|
| 161 |
+
raise ValueError(f"Unknown unquantized MoE backend: {backend.value}")
|
| 162 |
+
|
| 163 |
+
|
| 164 |
+
def map_unquantized_backend(runner_backend: MoEBackend) -> UnquantizedMoeBackend:
|
| 165 |
+
"""Map user's MoEBackend to UnquantizedMoeBackend."""
|
| 166 |
+
mapping = {
|
| 167 |
+
"triton": UnquantizedMoeBackend.TRITON,
|
| 168 |
+
"batched_triton": UnquantizedMoeBackend.BATCHED_TRITON,
|
| 169 |
+
"flashinfer_trtllm": UnquantizedMoeBackend.FLASHINFER_TRTLLM,
|
| 170 |
+
"flashinfer_cutlass": UnquantizedMoeBackend.FLASHINFER_CUTLASS,
|
| 171 |
+
"aiter": UnquantizedMoeBackend.AITER,
|
| 172 |
+
}
|
| 173 |
+
if backend := mapping.get(runner_backend):
|
| 174 |
+
return backend
|
| 175 |
+
raise ValueError(
|
| 176 |
+
f"moe_backend='{runner_backend}' is not supported for unquantized MoE. "
|
| 177 |
+
f"Expected one of {list(mapping.keys())}."
|
| 178 |
+
)
|
| 179 |
+
|
| 180 |
+
|
| 181 |
+
def _trtllm_bf16_lora_supported(moe_config: FusedMoEConfig) -> bool:
|
| 182 |
+
"""Gate for routing LoRA-enabled BF16 MoE to the FlashInfer TRT-LLM
|
| 183 |
+
gemm1_lora_delta path (PR #3153). Conservative: device + routing method;
|
| 184 |
+
the experts class's own _supports_* checks and the modular_kernel LoRA
|
| 185 |
+
gate provide the final filtering.
|
| 186 |
+
"""
|
| 187 |
+
from vllm.model_executor.layers.fused_moe.experts.trtllm_lora_moe import (
|
| 188 |
+
TrtLlmBf16LoRAExperts,
|
| 189 |
+
)
|
| 190 |
+
|
| 191 |
+
if not TrtLlmBf16LoRAExperts._supports_current_device():
|
| 192 |
+
return False
|
| 193 |
+
if not TrtLlmBf16LoRAExperts._supports_routing_method(
|
| 194 |
+
moe_config.routing_method, None, None
|
| 195 |
+
):
|
| 196 |
+
return False
|
| 197 |
+
if not TrtLlmBf16LoRAExperts._supports_parallel_config(
|
| 198 |
+
moe_config.moe_parallel_config
|
| 199 |
+
):
|
| 200 |
+
return False
|
| 201 |
+
# The flashinfer trtllm fused-MoE kernel requires the per-partition
|
| 202 |
+
# intermediate size to be a multiple of 128. Plain TP shards the MoE
|
| 203 |
+
# intermediate dim (e.g. 768 -> 192 at tp=4), which would crash the kernel
|
| 204 |
+
# at runtime; fall back to Triton in that case.
|
| 205 |
+
return moe_config.intermediate_size_per_partition % 128 == 0
|
| 206 |
+
|
| 207 |
+
|
| 208 |
+
def select_unquantized_moe_backend(
|
| 209 |
+
moe_config: FusedMoEConfig,
|
| 210 |
+
) -> tuple[UnquantizedMoeBackend, type[mk.FusedMoEExperts] | None]:
|
| 211 |
+
"""
|
| 212 |
+
Select the primary Unquantized MoE backend.
|
| 213 |
+
Note: Shape-specific fallbacks may still occur at runtime.
|
| 214 |
+
"""
|
| 215 |
+
|
| 216 |
+
if current_platform.is_tpu():
|
| 217 |
+
return UnquantizedMoeBackend.TPU, None
|
| 218 |
+
|
| 219 |
+
if current_platform.is_out_of_tree():
|
| 220 |
+
return UnquantizedMoeBackend.OOT, None
|
| 221 |
+
|
| 222 |
+
if moe_config.is_lora_enabled:
|
| 223 |
+
if _trtllm_bf16_lora_supported(moe_config):
|
| 224 |
+
from vllm.model_executor.layers.fused_moe.experts.trtllm_lora_moe import (
|
| 225 |
+
TrtLlmBf16LoRAExperts,
|
| 226 |
+
)
|
| 227 |
+
|
| 228 |
+
logger.info_once(
|
| 229 |
+
"Using TrtLlmBf16LoRAExperts Unquantized MoE LoRA backend "
|
| 230 |
+
"(TrtLlmBf16LoRAExperts)."
|
| 231 |
+
)
|
| 232 |
+
return UnquantizedMoeBackend.FLASHINFER_TRTLLM, TrtLlmBf16LoRAExperts
|
| 233 |
+
logger.info_once("Using TRITON Unquantized MoE LoRA backend")
|
| 234 |
+
return UnquantizedMoeBackend.TRITON, backend_to_kernel_cls(
|
| 235 |
+
UnquantizedMoeBackend.TRITON
|
| 236 |
+
)[0]
|
| 237 |
+
|
| 238 |
+
# NOTE: the kernels are selected in the following order.
|
| 239 |
+
AVAILABLE_BACKENDS = _get_priority_backends(moe_config)
|
| 240 |
+
|
| 241 |
+
# NOTE(rob): We need to peak into the P/F selection to determine
|
| 242 |
+
# if we are using the batched or standard expert format, which
|
| 243 |
+
# if not ideal. Once we unify TP + DP/EP, we can select P/F first.
|
| 244 |
+
activation_format = (
|
| 245 |
+
mk.FusedMoEActivationFormat.BatchedExperts
|
| 246 |
+
if moe_config.moe_parallel_config.use_batched_activation_format
|
| 247 |
+
or moe_config.moe_backend == "batched_triton"
|
| 248 |
+
else mk.FusedMoEActivationFormat.Standard
|
| 249 |
+
)
|
| 250 |
+
|
| 251 |
+
def _make_log_backend(backend: UnquantizedMoeBackend) -> str:
|
| 252 |
+
available_strs = [b.value for b in AVAILABLE_BACKENDS]
|
| 253 |
+
return (
|
| 254 |
+
f"Using {backend.value} Unquantized MoE backend out "
|
| 255 |
+
f"of potential backends: {available_strs}."
|
| 256 |
+
)
|
| 257 |
+
|
| 258 |
+
def _make_log_unsupported(
|
| 259 |
+
backend: UnquantizedMoeBackend, reason: str | None
|
| 260 |
+
) -> str:
|
| 261 |
+
if reason:
|
| 262 |
+
return (
|
| 263 |
+
f"Unquantized MoE backend {backend.value} does not support the "
|
| 264 |
+
f"deployment configuration since {reason}."
|
| 265 |
+
)
|
| 266 |
+
return (
|
| 267 |
+
f"Unquantized MoE backend '{backend.value}' does not support the "
|
| 268 |
+
"deployment configuration."
|
| 269 |
+
)
|
| 270 |
+
|
| 271 |
+
def _return_or_raise(
|
| 272 |
+
backend: UnquantizedMoeBackend,
|
| 273 |
+
config: FusedMoEConfig,
|
| 274 |
+
activation_format: mk.FusedMoEActivationFormat,
|
| 275 |
+
) -> tuple[UnquantizedMoeBackend, type[mk.FusedMoEExperts] | None]:
|
| 276 |
+
reason = None
|
| 277 |
+
for k_cls in backend_to_kernel_cls(backend):
|
| 278 |
+
supported, reason = k_cls.is_supported_config(
|
| 279 |
+
k_cls, config, None, None, activation_format
|
| 280 |
+
)
|
| 281 |
+
if supported:
|
| 282 |
+
logger.info_once(_make_log_backend(backend))
|
| 283 |
+
return backend, k_cls
|
| 284 |
+
raise ValueError(_make_log_unsupported(backend, reason))
|
| 285 |
+
|
| 286 |
+
runner_backend = moe_config.moe_backend
|
| 287 |
+
# Marlin is likewise weight-only: it has no unquantized kernel at all. A model
|
| 288 |
+
# that needs a global marlin setting for its quantized experts (NVFP4A16 has no
|
| 289 |
+
# other supported MoE backend) but also carries an *unquantized* MoE — e.g. the
|
| 290 |
+
# BF16 MTP draft layer of a quantized checkpoint — would otherwise fail outright
|
| 291 |
+
# with "moe_backend='marlin' is not supported for unquantized MoE".
|
| 292 |
+
# Send that layer to Triton rather than auto: auto would pick FlashInfer TRT-LLM,
|
| 293 |
+
# whose grouped-routing kernel is warp-limited to 32 experts per group, and models
|
| 294 |
+
# like Ling-3.0 (512 experts in 8 groups = 64 per group) either crash there or,
|
| 295 |
+
# worse, mis-route and produce a drafter the target never agrees with.
|
| 296 |
+
if runner_backend == "marlin":
|
| 297 |
+
runner_backend = "triton"
|
| 298 |
+
# 'humming' is quantization-only; an unquantized layer (e.g. excluded via
|
| 299 |
+
# modules_to_not_convert) falls through to auto instead of erroring.
|
| 300 |
+
if runner_backend not in ["auto", "humming"]:
|
| 301 |
+
requested_backend = map_unquantized_backend(runner_backend)
|
| 302 |
+
if (
|
| 303 |
+
activation_format == mk.FusedMoEActivationFormat.BatchedExperts
|
| 304 |
+
and requested_backend == UnquantizedMoeBackend.TRITON
|
| 305 |
+
):
|
| 306 |
+
requested_backend = UnquantizedMoeBackend.BATCHED_TRITON
|
| 307 |
+
|
| 308 |
+
return _return_or_raise(requested_backend, moe_config, activation_format)
|
| 309 |
+
|
| 310 |
+
# Handle explicit AITER FP8 configuration.
|
| 311 |
+
if envs.is_set("VLLM_ROCM_USE_AITER") or envs.is_set("VLLM_ROCM_USE_AITER_MOE"):
|
| 312 |
+
skip_aiter_moe = (
|
| 313 |
+
not envs.VLLM_ROCM_USE_AITER
|
| 314 |
+
or not envs.VLLM_ROCM_USE_AITER_MOE
|
| 315 |
+
or rocm_aiter_ops.is_rdna_aiter_enabled()
|
| 316 |
+
)
|
| 317 |
+
if skip_aiter_moe:
|
| 318 |
+
if UnquantizedMoeBackend.AITER in AVAILABLE_BACKENDS:
|
| 319 |
+
AVAILABLE_BACKENDS.remove(UnquantizedMoeBackend.AITER)
|
| 320 |
+
else:
|
| 321 |
+
backend = UnquantizedMoeBackend.AITER
|
| 322 |
+
return _return_or_raise(backend, moe_config, activation_format)
|
| 323 |
+
|
| 324 |
+
for backend in AVAILABLE_BACKENDS:
|
| 325 |
+
for k_cls in backend_to_kernel_cls(backend):
|
| 326 |
+
supported, reason = k_cls.is_supported_config(
|
| 327 |
+
k_cls, moe_config, None, None, activation_format
|
| 328 |
+
)
|
| 329 |
+
if supported:
|
| 330 |
+
logger.info_once(_make_log_backend(backend))
|
| 331 |
+
return backend, k_cls
|
| 332 |
+
|
| 333 |
+
logger.debug_once(_make_log_unsupported(backend, reason))
|
| 334 |
+
|
| 335 |
+
raise NotImplementedError(
|
| 336 |
+
"No Unquantized MoE backend supports the deployment configuration."
|
| 337 |
+
)
|
| 338 |
+
|
| 339 |
+
|
| 340 |
+
def convert_to_unquantized_kernel_format(
|
| 341 |
+
unquantized_backend: UnquantizedMoeBackend,
|
| 342 |
+
moe_config: FusedMoEConfig,
|
| 343 |
+
w13_weight: torch.Tensor,
|
| 344 |
+
w2_weight: torch.Tensor,
|
| 345 |
+
) -> tuple[torch.Tensor, torch.Tensor]:
|
| 346 |
+
if unquantized_backend == UnquantizedMoeBackend.AITER:
|
| 347 |
+
w13_weight, w2_weight = rocm_aiter_ops.shuffle_weights(w13_weight, w2_weight)
|
| 348 |
+
|
| 349 |
+
elif unquantized_backend == UnquantizedMoeBackend.FLASHINFER_CUTLASS:
|
| 350 |
+
if moe_config.is_act_and_mul:
|
| 351 |
+
# Swap halves to arrange as [w3; w1] (kernel expectation)
|
| 352 |
+
# Non-gated MoE: w13 is a single projection, no need to swap.
|
| 353 |
+
w13_weight = swap_w13_to_w31(w13_weight)
|
| 354 |
+
|
| 355 |
+
elif unquantized_backend == UnquantizedMoeBackend.FLASHINFER_TRTLLM:
|
| 356 |
+
is_act_and_mul = moe_config.is_act_and_mul
|
| 357 |
+
if not is_act_and_mul:
|
| 358 |
+
# Kernel requires intermediate_size_per_partition % 128 == 0 (BlockMajorK
|
| 359 |
+
# weight layout uses block_k=128). Pad along the intermediate dim when
|
| 360 |
+
# the model + TP split don't satisfy the constraint.
|
| 361 |
+
w13_weight, w2_weight, padded_intermediate = align_moe_weights_for_fi(
|
| 362 |
+
w13_weight, w2_weight, is_act_and_mul, min_alignment=128
|
| 363 |
+
)
|
| 364 |
+
moe_config.intermediate_size_per_partition = padded_intermediate
|
| 365 |
+
|
| 366 |
+
_cache_permute_indices: dict[torch.Size, torch.Tensor] = {}
|
| 367 |
+
w13_weight, w2_weight = convert_moe_weights_to_flashinfer_trtllm_block_layout(
|
| 368 |
+
_cache_permute_indices,
|
| 369 |
+
w13_weight,
|
| 370 |
+
w2_weight,
|
| 371 |
+
is_gated_act_gemm=is_act_and_mul,
|
| 372 |
+
)
|
| 373 |
+
|
| 374 |
+
if (
|
| 375 |
+
unquantized_backend == UnquantizedMoeBackend.TRITON
|
| 376 |
+
and current_platform.is_rocm()
|
| 377 |
+
and envs.VLLM_ROCM_MOE_PADDING
|
| 378 |
+
):
|
| 379 |
+
# Skip .contiguous(): it would undo the ROCm MoE weight padding.
|
| 380 |
+
return w13_weight, w2_weight
|
| 381 |
+
return w13_weight.contiguous(), w2_weight.contiguous()
|
| 382 |
+
|
| 383 |
+
|
| 384 |
+
def make_unquantized_moe_kernel(
|
| 385 |
+
quant_config: FusedMoEQuantConfig,
|
| 386 |
+
moe_config: FusedMoEConfig,
|
| 387 |
+
backend: UnquantizedMoeBackend,
|
| 388 |
+
experts_cls: type[mk.FusedMoEExperts],
|
| 389 |
+
routing_tables: tuple[torch.Tensor, torch.Tensor, torch.Tensor] | None = None,
|
| 390 |
+
) -> mk.FusedMoEKernel:
|
| 391 |
+
from vllm.model_executor.layers.fused_moe.utils import (
|
| 392 |
+
warn_if_moe_use_td_ineffective,
|
| 393 |
+
)
|
| 394 |
+
|
| 395 |
+
# Warn against the selected backend, not each probed candidate.
|
| 396 |
+
warn_if_moe_use_td_ineffective(backend.value, is_quantized=False)
|
| 397 |
+
|
| 398 |
+
# Create Prepare/Finalize
|
| 399 |
+
is_monolithic = issubclass(experts_cls, mk.FusedMoEExpertsMonolithic)
|
| 400 |
+
prepare_finalize = maybe_make_prepare_finalize(
|
| 401 |
+
moe=moe_config,
|
| 402 |
+
quant_config=quant_config,
|
| 403 |
+
routing_tables=routing_tables,
|
| 404 |
+
allow_new_interface=True,
|
| 405 |
+
use_monolithic=is_monolithic,
|
| 406 |
+
)
|
| 407 |
+
assert prepare_finalize is not None
|
| 408 |
+
|
| 409 |
+
logger.info_once("Using %s", prepare_finalize.__class__.__name__)
|
| 410 |
+
logger.info_once("Using %s MoE backend", experts_cls.__name__)
|
| 411 |
+
|
| 412 |
+
# Create Experts
|
| 413 |
+
if prepare_finalize.activation_format == mk.FusedMoEActivationFormat.BatchedExperts:
|
| 414 |
+
max_num_tokens = prepare_finalize.max_num_tokens_per_rank()
|
| 415 |
+
assert max_num_tokens is not None
|
| 416 |
+
experts = experts_cls(
|
| 417 |
+
moe_config=moe_config,
|
| 418 |
+
quant_config=quant_config,
|
| 419 |
+
max_num_tokens=max_num_tokens,
|
| 420 |
+
num_dispatchers=prepare_finalize.num_dispatchers(),
|
| 421 |
+
)
|
| 422 |
+
else:
|
| 423 |
+
experts = experts_cls(
|
| 424 |
+
moe_config=moe_config,
|
| 425 |
+
quant_config=quant_config,
|
| 426 |
+
)
|
| 427 |
+
|
| 428 |
+
kernel = mk.FusedMoEKernel(
|
| 429 |
+
prepare_finalize,
|
| 430 |
+
experts,
|
| 431 |
+
)
|
| 432 |
+
|
| 433 |
+
return kernel
|
| 434 |
+
|
| 435 |
+
|
| 436 |
+
# ---------------------------------------------------------------------------
|
| 437 |
+
# Class-based view (first PR of the #37753 series; see oracle/base.py).
|
| 438 |
+
# Methods delegate to the module-level functions above so behaviour is
|
| 439 |
+
# bit-identical with pre-class code.
|
| 440 |
+
# ---------------------------------------------------------------------------
|
| 441 |
+
|
| 442 |
+
|
| 443 |
+
class UnquantizedMoEKernelOracle(MoEKernelOracle[UnquantizedMoeBackend]):
|
| 444 |
+
"""Class-based view of the unquantized MoE kernel oracle.
|
| 445 |
+
|
| 446 |
+
Each method delegates to its module-level counterpart so that
|
| 447 |
+
instantiating and calling this class is bit-identical to calling
|
| 448 |
+
the standalone functions. Follow-up PRs may move logic from the
|
| 449 |
+
module-level functions into these methods.
|
| 450 |
+
"""
|
| 451 |
+
|
| 452 |
+
def backend_enum_cls(self) -> type[UnquantizedMoeBackend]:
|
| 453 |
+
return UnquantizedMoeBackend
|
| 454 |
+
|
| 455 |
+
def get_priority_backends(
|
| 456 |
+
self, moe_config: FusedMoEConfig
|
| 457 |
+
) -> list[UnquantizedMoeBackend]:
|
| 458 |
+
return _get_priority_backends(moe_config)
|
| 459 |
+
|
| 460 |
+
def backend_to_kernel_cls(
|
| 461 |
+
self, backend: UnquantizedMoeBackend
|
| 462 |
+
) -> list[type[mk.FusedMoEExperts]]:
|
| 463 |
+
return backend_to_kernel_cls(backend)
|
| 464 |
+
|
| 465 |
+
def map_backend(self, runner_backend: MoEBackend) -> UnquantizedMoeBackend:
|
| 466 |
+
return map_unquantized_backend(runner_backend)
|
| 467 |
+
|
| 468 |
+
def select_backend(
|
| 469 |
+
self,
|
| 470 |
+
moe_config: FusedMoEConfig,
|
| 471 |
+
weight_key: "QuantKey | None" = None,
|
| 472 |
+
activation_key: "QuantKey | None" = None,
|
| 473 |
+
) -> tuple[UnquantizedMoeBackend, type[mk.FusedMoEExperts] | None]:
|
| 474 |
+
assert weight_key is None and activation_key is None, (
|
| 475 |
+
"Weights and activations will never be quantized for "
|
| 476 |
+
"UnquantizedMoEKernelOracle"
|
| 477 |
+
)
|
| 478 |
+
return select_unquantized_moe_backend(moe_config)
|
| 479 |
+
|
| 480 |
+
def convert_to_kernel_format(
|
| 481 |
+
self,
|
| 482 |
+
backend: UnquantizedMoeBackend,
|
| 483 |
+
moe_config: FusedMoEConfig,
|
| 484 |
+
w13_weight: torch.Tensor,
|
| 485 |
+
w2_weight: torch.Tensor,
|
| 486 |
+
) -> tuple[torch.Tensor, torch.Tensor]:
|
| 487 |
+
return convert_to_unquantized_kernel_format(
|
| 488 |
+
backend, moe_config, w13_weight, w2_weight
|
| 489 |
+
)
|
| 490 |
+
|
| 491 |
+
def make_kernel(
|
| 492 |
+
self,
|
| 493 |
+
quant_config: FusedMoEQuantConfig,
|
| 494 |
+
moe_config: FusedMoEConfig,
|
| 495 |
+
backend: UnquantizedMoeBackend,
|
| 496 |
+
experts_cls: type[mk.FusedMoEExperts],
|
| 497 |
+
routing_tables: tuple[torch.Tensor, torch.Tensor, torch.Tensor] | None = None,
|
| 498 |
+
) -> mk.FusedMoEKernel:
|
| 499 |
+
return make_unquantized_moe_kernel(
|
| 500 |
+
quant_config, moe_config, backend, experts_cls, routing_tables
|
| 501 |
+
)
|