biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[Bugfix] Fix NVFP4 TRTLLM MoE non-gated support; add gsm8k for Nemotron-3-Nano FP8+NVFP4 (#34725)

Browse Source 
Signed-off-by: mgoin <mgoin64@gmail.com>
This commit is contained in:
Michael Goin
2026-02-18 12:39:22 -05:00
committed by GitHub
parent 6b3166a7c7
commit caeb887bf6
4 changed files with 20 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
tests/evals/gsm8k/configs/moe-refactor/Nemotron-Nano-30B-Fp8-ModelOpt-fi-trtllm.yaml Normal file
View File

@@ -0,0 +1,8 @@
model_name: "nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-FP8"
accuracy_threshold: 0.29
num_questions: 1319
num_fewshot: 5
server_args: "--enforce-eager --max-model-len 8192 --tensor-parallel-size 2"
env:
VLLM_USE_FLASHINFER_MOE_FP8: "1"
VLLM_FLASHINFER_MOE_BACKEND: "latency"

8
tests/evals/gsm8k/configs/moe-refactor/Nemotron-Nano-30B-NvFp4-ModelOpt-fi-cutlass.yaml Normal file
View File

@@ -0,0 +1,8 @@
model_name: "nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-NVFP4"
accuracy_threshold: 0.29
num_questions: 1319
num_fewshot: 5
server_args: "--enforce-eager --max-model-len 8192 --tensor-parallel-size 2"
env:
VLLM_USE_FLASHINFER_MOE_FP4: "1"
VLLM_FLASHINFER_MOE_BACKEND: "throughput"

2
tests/evals/gsm8k/configs/moe-refactor/config-b200.txt
View File

@@ -13,3 +13,5 @@ Llama-4-Scout-BF16-fi-cutlass.yaml
Llama-4-Scout-BF16-triton.yaml
Mixtral-8x7B-BF16-fi-cutlass.yaml
Mixtral-8x7B-BF16-triton.yaml
Nemotron-Nano-30B-Fp8-ModelOpt-fi-trtllm.yaml
Nemotron-Nano-30B-NvFp4-ModelOpt-fi-cutlass.yaml

2
vllm/model_executor/layers/quantization/utils/flashinfer_fp4_moe.py
View File

@@ -122,6 +122,8 @@ def is_supported_config_trtllm(
return False, _make_reason("routing method")
elif activation_format != mk.FusedMoEActivationFormat.Standard:
return False, _make_reason("activation format")
elif moe_config.hidden_dim % 512 != 0:
return False, _make_reason("hidden_dim must be divisible by 512")
return True, None