From 319d52aae036b3214b035b87d04b1c5a723e8c78 Mon Sep 17 00:00:00 2001 From: woodchenwu Date: Wed, 27 Aug 2025 19:12:14 +0800 Subject: [PATCH 1/7] add qwen2.5 code 14B dynamic config --- .../qwen2_5-14b_instruct_fp8_dynamic.yaml | 26 +++++++++++++++++++ 1 file changed, 26 insertions(+) create mode 100644 configs/qwen2_5/fp8_dynamic/qwen2_5-14b_instruct_fp8_dynamic.yaml diff --git a/configs/qwen2_5/fp8_dynamic/qwen2_5-14b_instruct_fp8_dynamic.yaml b/configs/qwen2_5/fp8_dynamic/qwen2_5-14b_instruct_fp8_dynamic.yaml new file mode 100644 index 00000000..6965f756 --- /dev/null +++ b/configs/qwen2_5/fp8_dynamic/qwen2_5-14b_instruct_fp8_dynamic.yaml @@ -0,0 +1,26 @@ +# Global configuration of pipeline +global: + save_path: ./output + +# Simplified Configuration for LLM compression +model: + name: Qwen + model_path: Qwen/Qwen2.5-7B-Instruct + trust_remote_code: true + low_cpu_mem_usage: true + use_cache: false + torch_dtype: auto + device_map: auto + +# Compression configuration +compression: + name: PTQ + quantization: + name: fp8_dynamic # Supported: fp8_static, fp8_dynamic, int4_awq, int4_gptq + bits: 8 # Quantization bits (4/8) + quant_method: + weight: "per-tensor" + activation: "per-tensor" + ignore_layers: # Skip quantization for these layers + - "lm_head" + - "model.embed_tokens" From ad8a31fdf631beb1adf82598e4c82824d5a9265c Mon Sep 17 00:00:00 2001 From: woodchenwu Date: Wed, 27 Aug 2025 19:26:48 +0800 Subject: [PATCH 2/7] add qwen2.5 code 14B dynamic config --- .../qwen2_5/fp8_dynamic/qwen2_5-14b_instruct_fp8_dynamic.yaml | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/configs/qwen2_5/fp8_dynamic/qwen2_5-14b_instruct_fp8_dynamic.yaml b/configs/qwen2_5/fp8_dynamic/qwen2_5-14b_instruct_fp8_dynamic.yaml index 6965f756..ab78cbfa 100644 --- a/configs/qwen2_5/fp8_dynamic/qwen2_5-14b_instruct_fp8_dynamic.yaml +++ b/configs/qwen2_5/fp8_dynamic/qwen2_5-14b_instruct_fp8_dynamic.yaml @@ -5,7 +5,7 @@ global: # Simplified Configuration for LLM compression model: name: Qwen - model_path: Qwen/Qwen2.5-7B-Instruct + model_path: Qwen/Qwen2.5-Coder-14B-Instruct trust_remote_code: true low_cpu_mem_usage: true use_cache: false From 40dcba8aafb7e24f99b5db079efabdfb31100176 Mon Sep 17 00:00:00 2001 From: woodchenwu Date: Wed, 27 Aug 2025 19:51:17 +0800 Subject: [PATCH 3/7] add qwen2.5 code 14B dynamic config --- .../qwen2_5/fp8_dynamic/qwen2_5-14b_instruct_fp8_dynamic.yaml | 2 +- configs/qwen2_5/fp8_static/qwen2_5-14b_instruct_fp8_static.yaml | 2 +- 2 files changed, 2 insertions(+), 2 deletions(-) diff --git a/configs/qwen2_5/fp8_dynamic/qwen2_5-14b_instruct_fp8_dynamic.yaml b/configs/qwen2_5/fp8_dynamic/qwen2_5-14b_instruct_fp8_dynamic.yaml index ab78cbfa..376b803b 100644 --- a/configs/qwen2_5/fp8_dynamic/qwen2_5-14b_instruct_fp8_dynamic.yaml +++ b/configs/qwen2_5/fp8_dynamic/qwen2_5-14b_instruct_fp8_dynamic.yaml @@ -5,7 +5,7 @@ global: # Simplified Configuration for LLM compression model: name: Qwen - model_path: Qwen/Qwen2.5-Coder-14B-Instruct + model_path: Qwen/Qwen2.5-14B-Instruct trust_remote_code: true low_cpu_mem_usage: true use_cache: false diff --git a/configs/qwen2_5/fp8_static/qwen2_5-14b_instruct_fp8_static.yaml b/configs/qwen2_5/fp8_static/qwen2_5-14b_instruct_fp8_static.yaml index 48bf2900..7353bc93 100644 --- a/configs/qwen2_5/fp8_static/qwen2_5-14b_instruct_fp8_static.yaml +++ b/configs/qwen2_5/fp8_static/qwen2_5-14b_instruct_fp8_static.yaml @@ -5,7 +5,7 @@ global: # Simplified Configuration for LLM compression model: name: Qwen - model_path: Qwen/Qwen2.5-Coder-14B-Instruct + model_path: Qwen/Qwen2.5-14B-Instruct trust_remote_code: true low_cpu_mem_usage: true use_cache: false From 273898a439146fa73e0df832b1e56968b4ffceaf Mon Sep 17 00:00:00 2001 From: woodchenwu Date: Thu, 26 Feb 2026 11:29:53 +0800 Subject: [PATCH 4/7] update arxiv link --- README.md | 2 +- README_cn.md | 2 +- 2 files changed, 2 insertions(+), 2 deletions(-) diff --git a/README.md b/README.md index 9d5e334a..062ec6cb 100644 --- a/README.md +++ b/README.md @@ -12,7 +12,7 @@ A more accessible, comprehensive, and efficient toolkit for large model compress

- ✒️ TechnicalReport   |    📖 Documentation   |   🤗 Hugging Face   |   🤖 ModelScope + ✒️ TechnicalReport   |    📖 Documentation   |   🤗 Hugging Face   |   🤖 ModelScope

diff --git a/README_cn.md b/README_cn.md index 4520b188..455aa420 100644 --- a/README_cn.md +++ b/README_cn.md @@ -12,7 +12,7 @@

- ✒️ TechnicalReport   |    📖 Documentation   |   🤗 Hugging Face   |   🤖 ModelScope + ✒️ TechnicalReport   |    📖 Documentation   |   🤗 Hugging Face   |   🤖 ModelScope

From 26e31dfced80640c5a8b52f09aa82d32ff038d3a Mon Sep 17 00:00:00 2001 From: woodchenwu Date: Thu, 26 Feb 2026 12:41:00 +0800 Subject: [PATCH 5/7] update citation --- README.md | 11 +++++------ README_cn.md | 11 +++++------ 2 files changed, 10 insertions(+), 12 deletions(-) diff --git a/README.md b/README.md index 062ec6cb..fdd08fe5 100644 --- a/README.md +++ b/README.md @@ -1031,12 +1031,11 @@ The code for this project is open-sourced under the [License for AngelSlim](LICE ## 🔗 Citation ``` -@software{AngelSlim2025, - title={{AngelSlim}}, - author={Tencent AngelSlim Project Contributors}, - year={2025}, - month={6}, - url={https://github.com/Tencent/AngelSlim}, +@article{angelslim2026, + title={AngelSlim: A more accessible, comprehensive, and efficient toolkit for large model compression}, + author={Hunyuan AI Infra Team}, + journal={arXiv preprint arXiv:2602.21233}, + year={2026} } ``` diff --git a/README_cn.md b/README_cn.md index 455aa420..eb759605 100644 --- a/README_cn.md +++ b/README_cn.md @@ -1031,12 +1031,11 @@ Qwen3-Omni系列模型的`BF16`、`FP8-Static`、`FP8-Dynamic`在`aime25`、`gpq ## 🔗引用 ``` -@software{AngelSlim2025, - title={{AngelSlim}}, - author={Tencent AngelSlim Project Contributors}, - year={2025}, - month={7}, - url={https://github.com/Tencent/AngelSlim}, +@article{angelslim2026, + title={AngelSlim: A more accessible, comprehensive, and efficient toolkit for large model compression}, + author={Hunyuan AI Infra Team}, + journal={arXiv preprint arXiv:2602.21233}, + year={2026} } ``` From 626a3c9f957808c5f93543952a49249e2a292bab Mon Sep 17 00:00:00 2001 From: woodchenwu Date: Thu, 26 Feb 2026 15:45:40 +0800 Subject: [PATCH 6/7] update https --- README.md | 2 +- README_cn.md | 2 +- 2 files changed, 2 insertions(+), 2 deletions(-) diff --git a/README.md b/README.md index fdd08fe5..11ef363e 100644 --- a/README.md +++ b/README.md @@ -12,7 +12,7 @@ A more accessible, comprehensive, and efficient toolkit for large model compress

- ✒️ TechnicalReport   |    📖 Documentation   |   🤗 Hugging Face   |   🤖 ModelScope + ✒️ TechnicalReport   |    📖 Documentation   |   🤗 Hugging Face   |   🤖 ModelScope

diff --git a/README_cn.md b/README_cn.md index eb759605..658e78f3 100644 --- a/README_cn.md +++ b/README_cn.md @@ -12,7 +12,7 @@

- ✒️ TechnicalReport   |    📖 Documentation   |   🤗 Hugging Face   |   🤖 ModelScope + ✒️ TechnicalReport   |    📖 Documentation   |   🤗 Hugging Face   |   🤖 ModelScope

From a1b394b50d9ae0e2a5abee54172592783240bb2e Mon Sep 17 00:00:00 2001 From: woodchenwu Date: Mon, 6 Jul 2026 13:58:09 +0800 Subject: [PATCH 7/7] update readme news for HY3.0 --- README.md | 1 + README_cn.md | 1 + 2 files changed, 2 insertions(+) diff --git a/README.md b/README.md index f40ddf55..48d53395 100644 --- a/README.md +++ b/README.md @@ -22,6 +22,7 @@ A more accessible, comprehensive, and efficient toolkit for large model compress

## 📣Latest News +- [26/07/06] We now support FP8-Static quantization , SmoothQuant for **Hy3** (MoE A20B).[[Docs]](https://github.com/Tencent/AngelSlim/tree/main/scripts/ptq) - [26/06/04] We have released **Stem**, a sparse attention algorithm that accelerates the **Prefill** stage of long-context LLMs by dynamically selecting top-k key blocks for block-sparse attention, significantly reducing latency while preserving generation quality. [[Docs]](https://angelslim.readthedocs.io/zh-cn/latest/features/sparse_attention/stem.html) - [26/06/01] We have released **DFlare**, a block-diffusion speculative decoding framework with layer-wise fusion that achieves up to **5.52× end-to-end speedup**. [[Docs]](https://angelslim.readthedocs.io/zh-cn/latest/features/speculative_decoding/dflare.html) - [26/05/27] We have released **D-Cut**, an adaptive verification depth pruning technique for speculative decoding. [[Docs]](https://angelslim.readthedocs.io/zh-cn/latest/dcut.html) diff --git a/README_cn.md b/README_cn.md index 0e326829..92af9a72 100644 --- a/README_cn.md +++ b/README_cn.md @@ -22,6 +22,7 @@

## 📣最新进展 +- [26/07/06] 我们支持了 **Hy3**（MoE A20B）模型的 FP8-Static 量化，SmoothQuant 量化. [[文档]](https://github.com/Tencent/AngelSlim/tree/main/scripts/ptq) - [26/06/04] 我们发布了 **Stem**，一种稀疏注意力算法，通过在 block 粒度动态选择 top-k 关键块执行 block-sparse attention，加速长上下文 LLM 的 **Prefill** 阶段，在大幅降低延迟的同时实现几乎无损的生成质量。[[文档]](https://angelslim.readthedocs.io/zh-cn/latest/features/sparse_attention/stem.html) - [26/06/01] 我们发布了 **DFlare**，一种基于 layer-wise fusion 的块扩散投机解码框架，端到端加速比可达 **5.52×**。[[文档]](https://angelslim.readthedocs.io/zh-cn/latest/features/speculative_decoding/dflare.html) - [26/05/27] 我们发布了 **D-Cut**，一种用于投机解码的自适应验证深度裁剪技术。[[文档]](https://angelslim.readthedocs.io/zh-cn/latest/dcut.html)