A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 3, 06:45 UTC
[CI] Add DSV4-Flash to gsm8k moe-refactor/config-b200.txt (#42111)
Signed-off-by: mgoin <mgoin64@gmail.com>
Michael Goin committed
May 19, 2026 at 23:21 UTC
cd0ff26e7acf2c691a33d4c44276db6980bab24b
6 files changed
+12
-1
requirements/common.txt
+1
@@ -9,6 +9,7 @@ blake3
9
py-cpuinfo
10
transformers >= 4.56.0, != 5.0.*, != 5.1.*, != 5.2.*, != 5.3.*, != 5.4.*, != 5.5.0
11
tokenizers >= 0.21.1 # Required for fast incremental detokenization.
12
+safetensors >= 0.6.2 # MXFP4/MXFP6 dtype support (F8_E8M0, F4) added in 0.6.0: https://github.com/huggingface/safetensors/pull/611
13
protobuf >= 5.29.6, !=6.30.*, !=6.31.*, !=6.32.*, !=6.33.0.*, !=6.33.1.*, !=6.33.2.*, !=6.33.3.*, !=6.33.4.* # Required by LlamaTokenizer, gRPC. CVE-2026-0994
14
fastapi[standard] >= 0.115.0 # Required by FastAPI's form models in the OpenAI API server's audio transcriptions endpoint.
15
aiohttp >= 3.13.3
requirements/test/cuda.txt
+2
-1
@@ -858,8 +858,9 @@ s3transfer==0.10.3
858
# via boto3
859
sacrebleu==2.4.3
860
# via lm-eval
861
-safetensors==0.4.5
861
+safetensors==0.7.0
862
# via
863
+ # -c requirements/common.txt
864
# accelerate
865
# open-clip-torch
866
# peft
requirements/test/rocm.txt
+2
@@ -1048,6 +1048,8 @@ sacrebleu==2.6.0
1048
# via lm-eval
1049
safetensors==0.7.0
1050
# via
1051
+ # -c requirements/common.txt
1052
+ # -r requirements/test/../common.txt
1053
# accelerate
1054
# open-clip-torch
1055
# peft
requirements/test/xpu.txt
+1
@@ -539,6 +539,7 @@ sacrebleu==2.6.0
539
# via lm-eval
540
safetensors==0.7.0
541
# via
542
+ # -c requirements/common.txt
543
# accelerate
544
# timm
545
# transformers
tests/evals/gsm8k/configs/moe-refactor/DeepSeek-V4-Flash-deepgemm-mega-moe.yaml
new
+5
@@ -0,0 +1,5 @@
1
+model_name: "deepseek-ai/DeepSeek-V4-Flash"
2
+accuracy_threshold: 0.95
3
+num_questions: 1319
4
+num_fewshot: 5
5
+server_args: "--trust-remote-code --kv-cache-dtype fp8 --block-size 256 --enable-expert-parallel --tensor-parallel-size 2 --attention_config.use_fp4_indexer_cache=True --moe-backend deep_gemm_mega_moe --tokenizer-mode deepseek_v4 --tool-call-parser deepseek_v4 --enable-auto-tool-choice --reasoning-parser deepseek_v4 --speculative_config.method=mtp --speculative_config.num_speculative_tokens=2"
tests/evals/gsm8k/configs/moe-refactor/config-b200.txt
+1
@@ -16,3 +16,4 @@ Mixtral-8x7B-BF16-triton.yaml
16
Nemotron-Nano-30B-Fp8-ModelOpt-fi-trtllm.yaml
17
Nemotron-Nano-30B-NvFp4-ModelOpt-fi-cutlass.yaml
18
Nemotron-Nano-30B-NvFp4-ModelOpt-vllm-cutlass.yaml
19
+DeepSeek-V4-Flash-deepgemm-mega-moe.yaml