SignRoundV2: Closing the Performance Gap in Extremely Low-Bit Post-Training Quantization for LLMs
Paper
•
2512.04746
•
Published
•
7
Need4Speed
company
AI & ML interests
None defined yet.
Recent Activity
View all activity
Haihao
authored
a
paper
1 day ago
wenhuach
authored
a
paper
1 day ago
Post
1603
🚀 SignRoundV2 for LLM quantization: PTQ-level cost, QAT-level accuracy — yes, even at 2 bits.
SignRoundV2: Closing the Performance Gap in Extremely Low-Bit Post-Training Quantization for LLMs (2512.04746)
SignRoundV2: Closing the Performance Gap in Extremely Low-Bit Post-Training Quantization for LLMs (2512.04746)
Post
297
🚀 AutoRound(https://github.com/intel/auto-round) is now supported by SGLang!
After integrations with TorchAO, Transformers, and VLLM, AutoRound-quantized models are now officially compatible with SGLang — bringing faster and more flexible deployment to your LLM workflows.
💡 We’ve also enhanced the RTN mode (--iters 0), cutting quantization costs significantly for low-resource users.
⭐ Star our repo and stay tuned for more exciting updates!
After integrations with TorchAO, Transformers, and VLLM, AutoRound-quantized models are now officially compatible with SGLang — bringing faster and more flexible deployment to your LLM workflows.
💡 We’ve also enhanced the RTN mode (--iters 0), cutting quantization costs significantly for low-resource users.
⭐ Star our repo and stay tuned for more exciting updates!
Post
1758
AutoRound keeps evolving its LLM quantization algorithm! 🚀
After enhancing W2A16 quantization, we now offer a fast algorithm to generate mixed bits/data-type schemes (~2mins for 8B models), great for MXFP4 and W2A16.
Learn more: https://github.com/intel/auto-round/blob/main/docs/step_by_step.md#autoscheme
After enhancing W2A16 quantization, we now offer a fast algorithm to generate mixed bits/data-type schemes (~2mins for 8B models), great for MXFP4 and W2A16.
Learn more: https://github.com/intel/auto-round/blob/main/docs/step_by_step.md#autoscheme
lvwerra
authored
a
paper
about 2 months ago
Post
421
AutoRound v0.7 is out! 🚀
This release includes enhanced algorithms for W2A16, NVFP4, and MXFP4, along with support for FP8 models as input.
👉 Check out the full details here: https://github.com/intel/auto-round/releases/tag/v0.7.0
This release includes enhanced algorithms for W2A16, NVFP4, and MXFP4, along with support for FP8 models as input.
👉 Check out the full details here: https://github.com/intel/auto-round/releases/tag/v0.7.0
Post
1941
🚀 AutoRound(https://github.com/intel/auto-round) Now Supports GGUF Export & Custom Bit Settings!
We're excited to announce that AutoRound now supports:
✅ GGUF format export – for seamless compatibility with popular inference engines.
✅ Custom bit settings – tailor quantization to your needs for optimal performance.
Check out these newly released models:
🔹Intel/Qwen3-235B-A22B-Instruct-2507-gguf-q4km-AutoRound
🔹Intel/Qwen3-235B-A22B-Instruct-2507-gguf-q2ks-mixed-AutoRound
🔹Intel/Kimi-K2-Instruct-gguf-q2ks-mixed-AutoRound
Stay tuned! An even more advanced algorithm for some configurations is coming soon.
We're excited to announce that AutoRound now supports:
✅ GGUF format export – for seamless compatibility with popular inference engines.
✅ Custom bit settings – tailor quantization to your needs for optimal performance.
Check out these newly released models:
🔹Intel/Qwen3-235B-A22B-Instruct-2507-gguf-q4km-AutoRound
🔹Intel/Qwen3-235B-A22B-Instruct-2507-gguf-q2ks-mixed-AutoRound
🔹Intel/Kimi-K2-Instruct-gguf-q2ks-mixed-AutoRound
Stay tuned! An even more advanced algorithm for some configurations is coming soon.
lvwerra
authored
a
paper
5 months ago
loubnabnl
authored
a
paper
6 months ago
zhentaoyu
authored
a
paper
6 months ago
Post
1909
AutoRound(https://github.com/intel/auto-round) has been integrated into vLLM , allowing you to run AutoRound-formatted models directly in the upcoming release.
Beside, we strongly recommend using AutoRound to generate AWQ INT4 models, as AutoAWQ is no longer maintained and manually configuring new models is not trivial due to the need for custom layer mappings.
Beside, we strongly recommend using AutoRound to generate AWQ INT4 models, as AutoAWQ is no longer maintained and manually configuring new models is not trivial due to the need for custom layer mappings.
Post
5817
SmolVLM is now available on PocketPal — you can run it offline on your smartphone to interpret the world around you. 🌍📱
And check out this real-time camera demo by @ngxson , powered by llama.cpp:
https://github.com/ngxson/smolvlm-realtime-webcam
https://x.com/pocketpal_ai
And check out this real-time camera demo by @ngxson , powered by llama.cpp:
https://github.com/ngxson/smolvlm-realtime-webcam
https://x.com/pocketpal_ai
-
- 4 replies
zhentaoyu
authored
a
paper
7 months ago
Post
1941
AutoRound(https://github.com/intel/auto-round) has been integrated into Transformers, allowing you to run AutoRound-formatted models directly in the upcoming release. Additionally, we are actively working on supporting the GGUF double-quant format, e.g. q4_k_s, stay tuned!
https://huggingface.co/blog/autoround
https://huggingface.co/blog/autoround
lvwerra
authored
a
paper
8 months ago
loubnabnl
authored
a
paper
8 months ago
Haihao
authored
a
paper
9 months ago
Post
2537
Check out [DeepSeek-R1 INT2 model(
OPEA/DeepSeek-R1-int2-mixed-sym-inc). This 200GB DeepSeek-R1 model shows only about a 2% drop in MMLU, though it's quite slow due to kernel issue.
| | BF16 | INT2-mixed |
| ------------- | ------ | ---------- |
| mmlu | 0.8514 | 0.8302 |
| hellaswag | 0.6935 | 0.6657 |
| winogrande | 0.7932 | 0.7940 |
| arc_challenge | 0.6212 | 0.6084 |
| | BF16 | INT2-mixed |
| ------------- | ------ | ---------- |
| mmlu | 0.8514 | 0.8302 |
| hellaswag | 0.6935 | 0.6657 |
| winogrande | 0.7932 | 0.7940 |
| arc_challenge | 0.6212 | 0.6084 |
Post
748
OPEA Space has released several quantized DeepSeek models, including INT2. Explore them here
OPEA/deepseek-6784a012d91191015587584a
OPEA/deepseek-6784a012d91191015587584a