Support DeepSeek-V4 AMD Quark NVFP4 with emulation kernel (#47972)

Signed-off-by: Jimmy Adams <Adam_ji@outlook.com>
Signed-off-by: Andreas Karatzas <akaratza@amd.com>
Signed-off-by: huji <huidong.ji@amd.com>
Signed-off-by: jimmy-adams <41593649+jimmy-adams@users.noreply.github.com>
Co-authored-by: Andreas Karatzas <akaratza@amd.com>
Co-authored-by: huji <huidong.ji@amd.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
This commit is contained in:
jimmy-adams
2026-08-07 01:19:31 -07:00
committed by GitHub
co-authored by Andreas Karatzas huji Cursor
parent 0de0362ea1
commit da788334bc
11 changed files with 332 additions and 17 deletions
@@ -0,0 +1,15 @@
model_name: "amd/DeepSeek-V4-Flash-NVFP4"
accuracy_threshold: 0.92
num_questions: 1319
num_fewshot: 8
startup_max_wait_seconds: 1800
server_args: >-
--tensor-parallel-size 8
--max-model-len 4096
--trust-remote-code
--kv-cache-dtype fp8
--tokenizer-mode deepseek_v4
--compilation-config '{"mode": 3, "cudagraph_mode": "FULL_DECODE_ONLY"}'
env:
VLLM_ROCM_USE_AITER: "1"
VLLM_ROCM_USE_AITER_MOE: "1"
@@ -0,0 +1,15 @@
model_name: "amd/DeepSeek-V4-Pro-NVFP4"
accuracy_threshold: 0.92
num_questions: 1319
num_fewshot: 8
startup_max_wait_seconds: 3600
server_args: >-
--tensor-parallel-size 8
--max-model-len 4096
--trust-remote-code
--kv-cache-dtype fp8
--tokenizer-mode deepseek_v4
--compilation-config '{"mode": 3, "cudagraph_mode": "FULL_DECODE_ONLY"}'
env:
VLLM_ROCM_USE_AITER: "1"
VLLM_ROCM_USE_AITER_MOE: "1"
@@ -0,0 +1,8 @@
# GFX950 model configurations for GSM8K evaluation
# TP8 models only
# DeepSeek-V4-Pro-NVFP4.yaml is temporarily disabled: repeated CI failures due to
# cold NFS reads of the 850GiB checkpoint exceeding startup_max_wait_seconds
# (see https://github.com/vllm-project/vllm/pull/47972). Re-enable once the
# startup timeout fix has landed and is verified green.
# DeepSeek-V4-Pro-NVFP4.yaml
DeepSeek-V4-Flash-NVFP4.yaml