From c9f6b9ba259569d287c9dd75c2fce631a0b02fee Mon Sep 17 00:00:00 2001 From: amote-i <49533125+amote-i@users.noreply.github.com> Date: Sun, 23 Aug 2026 13:05:03 +0800 Subject: [PATCH] [NPU] [DOC] Refresh supported features and models on Ascend NPU (#35836) --- .../reference/support_features.mdx | 90 +++++++++++----- .../ascend-npus/reference/support_models.mdx | 102 ++++++++++++++++-- 2 files changed, 159 insertions(+), 33 deletions(-) diff --git a/docs/docs/hardware-platforms/ascend-npus/reference/support_features.mdx b/docs/docs/hardware-platforms/ascend-npus/reference/support_features.mdx index 4c0fe9429..c200b1fd2 100644 --- a/docs/docs/hardware-platforms/ascend-npus/reference/support_features.mdx +++ b/docs/docs/hardware-platforms/ascend-npus/reference/support_features.mdx @@ -429,6 +429,12 @@ click [Server Arguments](../../../advanced_features/server_arguments). Type: int A2, A3 + + `--retraction-policy` + `length` + `length`,
`priority` + A2, A3 + `--schedule-conservativeness` `1.0` @@ -445,13 +451,13 @@ click [Server Arguments](../../../advanced_features/server_arguments). `--swa-full-tokens-ratio` `0.8` Type: float - Planned + A2, A3 `--disable-hybrid-swa-memory` `False` bool flag
(set to enable) - Planned + A2, A3 `--radix-eviction-policy` @@ -511,7 +517,7 @@ click [Server Arguments](../../../advanced_features/server_arguments). `--enable-dynamic-chunking` `False` bool flag
(set to enable) - Experimental + A2, A3 @@ -563,19 +569,19 @@ click [Server Arguments](../../../advanced_features/server_arguments). `--moe-data-parallel-size`
`--moe-dp-size` `1` Type: int - Planned + A2, A3 `--pp-max-micro-batch-size` `None` Type: int - Experimental + A2, A3 `--pp-async-batch-depth` - `None` + `0` Type: int - Experimental + A2, A3 `--stream-interval` @@ -607,6 +613,12 @@ click [Server Arguments](../../../advanced_features/server_arguments). bool flag (set to enable) A2, A3 + + `--enable-session-radix-cache` + `False` + bool flag
(set to enable) + A2, A3 + `--random-seed` `None` @@ -653,7 +665,7 @@ click [Server Arguments](../../../advanced_features/server_arguments). `--model-checksum` `None` Type: str - Planned + A2, A3 `--base-gpu-id` @@ -1010,6 +1022,12 @@ click [Server Arguments](../../../advanced_features/server_arguments). `deepseek-r1`
`deepseek-v3`
`glm45`
`gpt-oss`
`kimi`
`qwen3`
`qwen3-thinking`
`step3` A2, A3 + + `--default-chat-template-kwargs` + `None` + Type: JSON / Dict + A2, A3 + `--strip-thinking-cache` `False` @@ -1247,7 +1265,7 @@ click [Server Arguments](../../../advanced_features/server_arguments). `--lora-strict-loading` `False` Type: bool - Special for GPU + A2, A3 `--lora-drain-wait-threshold` @@ -1436,7 +1454,7 @@ click [Server Arguments](../../../advanced_features/server_arguments). `--speculative-token-map` `None` - Type: str + Type: str; requires `--tp-size=1` A2, A3 @@ -1514,49 +1532,49 @@ click [Server Arguments](../../../advanced_features/server_arguments). `--speculative-ngram-min-bfs-breadth` `1` Type: int - Experimental + A2, A3 `--speculative-ngram-max-bfs-breadth` `10` Type: int - Experimental + A2, A3 `--speculative-ngram-match-type` `BFS` `BFS`,
`PROB` - Experimental. `BFS` uses recency-based expansion; `PROB` uses frequency-based expansion. + A2, A3. `BFS` uses recency-based expansion; `PROB` uses frequency-based expansion. `--speculative-ngram-max-trie-depth` `18` Type: int - Experimental + A2, A3 `--speculative-ngram-capacity` `10000000` Type: int - Experimental + A2, A3 `--speculative-ngram-external-corpus-path` `None` Type: str - Experimental + A2, A3 `--speculative-ngram-external-sam-budget` `0` Type: int - Experimental + A2, A3 `--speculative-ngram-external-corpus-max-tokens` `10000000` Type: int - Experimental + A2, A3 `--speculative-adaptive` @@ -1615,7 +1633,7 @@ click [Server Arguments](../../../advanced_features/server_arguments). `--moe-runner-backend` `auto` `auto`, `triton` - Special for GPU + A2, A3 `--flashinfer-mxfp4-moe-precision` @@ -1635,6 +1653,12 @@ click [Server Arguments](../../../advanced_features/server_arguments). `normal`,
`low_latency`,
`auto` A2, A3 + + `--fuseep-mode` + `2` + `1`,
`2` + A2, A3 + `--deepep-dispatcher-output-dtype` `auto` @@ -2136,7 +2160,7 @@ If the value is int8, you must also set the environment variable:DEEP_NORMAL_MOD `--disable-custom-all-reduce` `False` bool flag
(set to enable) - Special for GPU + A2, A3 `--enable-mscclpp` @@ -2382,7 +2406,7 @@ If the value is int8, you must also set the environment variable:DEEP_NORMAL_MOD `--rl-on-policy-target` `None` `fsdp` - Planned + A2, A3 `--disable-attn-tp-gather` @@ -2400,7 +2424,7 @@ If the value is int8, you must also set the environment variable:DEEP_NORMAL_MOD `--enable-attn-tp-input-scattered` `False` bool flag
(set to enable) - Experimental + A2, A3 `--enable-prefill-cp` @@ -2698,25 +2722,25 @@ If the value is int8, you must also set the environment variable:DEEP_NORMAL_MOD `--remote-instance-weight-loader-seed-instance-ip` `None` Type: str - Special for GPU + A2, A3 `--remote-instance-weight-loader-seed-instance-service-port` `None` Type: int - Special for GPU + A2, A3 `--remote-instance-weight-loader-send-weights-group-ports` `None` Type: JSON
list - Special for GPU + A2, A3 `--remote-instance-weight-loader-backend` `nccl` - `transfer_engine`,
`nccl` - Special for GPU + `transfer_engine`,
`nccl`,
`modelexpress` + A2, A3 `--weight-loader-drop-cache-after-load` @@ -2811,6 +2835,18 @@ If the value is int8, you must also set the environment variable:DEEP_NORMAL_MOD bool flag
(set to enable) A2, A3 + + `--mm-processor-worker-num` + `0` + Type: int + A2, A3 + + + `--mm-io-worker-num` + `0` + Type: int + A2, A3 + `--limit-mm-data-per-request` `None` diff --git a/docs/docs/hardware-platforms/ascend-npus/reference/support_models.mdx b/docs/docs/hardware-platforms/ascend-npus/reference/support_models.mdx index 0ad1be268..32743c162 100644 --- a/docs/docs/hardware-platforms/ascend-npus/reference/support_models.mdx +++ b/docs/docs/hardware-platforms/ascend-npus/reference/support_models.mdx @@ -61,6 +61,12 @@ You are welcome to enable various models based on your business requirements. ✅ ✅ + + Qwen/Qwen3.8-2.4T-A95B + Qwen3.8 + ✅ + ✅ + Qwen/Qwen3.6-35B-A3B Qwen3.6 @@ -223,12 +229,6 @@ You are welcome to enable various models based on your business requirements. ✅ ✅ - - huihui-ai/grok-2 - Grok - ✅ - ✅ - ZhipuAI/chatglm2-6b ChatGLM @@ -277,6 +277,12 @@ You are welcome to enable various models based on your business requirements. ✅ ✅ + + iridiumine/MiMo-V2-Flash-W8A8 + MiMo + ✅ + ✅ + XiaomiMiMo/MiMo-7B-RL MiMo @@ -331,6 +337,12 @@ You are welcome to enable various models based on your business requirements. ✅ ✅ + + sgl-npu/Kimi-K3-W4A8 + Kimi + ✅ + ✅ + Eco-Tech/Kimi-K2.6-w4a8 Kimi @@ -409,6 +421,48 @@ You are welcome to enable various models based on your business requirements. ✅ ✅ + + LiquidAI/LFM2.5-1.2B-Instruct + LFM2 (230M, 350M, 1.2B) + ✅ + ✅ + + + LiquidAI/LFM2-8B-A1B + LFM2-MoE (8B-A1B, 24B-A2B) + ✅ + ✅ + + + tiiuae/Falcon-H1-34B-Instruct + Falcon-H1 (0.5B-34B) + ✅ + ✅ + + + ibm-granite/granite-4.0-h-micro + IBM Granite 4.0 (Hybrid, Dense) + ✅ + ✅ + + + ibm-granite/granite-4.0-micro + IBM Granite 4.0 (Hybrid, Dense) + ✅ + ✅ + + + poolside/Laguna-XS.2 + Laguna XS.2 (poolside) + ✅ + ✅ + + + Tencent-Hunyuan/Hy3 + Hunyuan + ✅ + ✅ + @@ -568,6 +622,42 @@ You are welcome to enable various models based on your business requirements. ✅ ✅ + + deepseek-ai/DeepSeek-OCR-2 + DeepSeek-OCR / OCR-2 + ✅ + ✅ + + + lmms-lab/LLaVA-NeXT-Video-7B + LLaVA video variants (LLaVA-NeXT-Video, LLaVA-OneVision) + ✅ + ✅ + + + ZhipuAI/GLM-OCR + GLM-OCR + ✅ + ✅ + + + AI-ModelScope/whisper-large-v3 + Whisper + ✅ + ✅ + + + Qwen/Qwen3-ASR-1.7B + Qwen3-ASR (0.6B, 1.7B) + ✅ + ✅ + + + LiquidAI/LFM2.5-VL-1.6B + LFM2-VL + ✅ + ✅ +