[CI] Raise GSM8K startup max wait to 2400s for flaky 30B+ MoE eval configs (#57361)
Signed-off-by: khluu <khluu000@gmail.com>
This commit is contained in:
@@ -3,6 +3,8 @@ accuracy_threshold: 0.90
|
|||||||
num_questions: 1319
|
num_questions: 1319
|
||||||
num_fewshot: 5
|
num_fewshot: 5
|
||||||
max_concurrency: 100
|
max_concurrency: 100
|
||||||
|
# 30B+ MoE with EP/PCP: download + kernel autotuning can exceed the 1200s default.
|
||||||
|
startup_max_wait_seconds: 2400
|
||||||
# Leave headroom for FlashInfer CUTLASS autotuning after KV-cache allocation.
|
# Leave headroom for FlashInfer CUTLASS autotuning after KV-cache allocation.
|
||||||
server_args: >-
|
server_args: >-
|
||||||
--compilation-config '{"cudagraph_mode":"NONE"}'
|
--compilation-config '{"cudagraph_mode":"NONE"}'
|
||||||
|
|||||||
+2
@@ -2,6 +2,8 @@ model_name: "Qwen/Qwen3-Coder-30B-A3B-Instruct-FP8"
|
|||||||
accuracy_threshold: 0.88
|
accuracy_threshold: 0.88
|
||||||
num_questions: 1319
|
num_questions: 1319
|
||||||
num_fewshot: 5
|
num_fewshot: 5
|
||||||
|
# 30B MoE + DeepGEMM/DeepEP JIT compile can exceed the 1200s default startup wait.
|
||||||
|
startup_max_wait_seconds: 2400
|
||||||
server_args: "--enforce-eager --max-model-len 8192 --data-parallel-size 2 --enable-expert-parallel --all2all-backend deepep_low_latency --disable-uvicorn-access-log"
|
server_args: "--enforce-eager --max-model-len 8192 --data-parallel-size 2 --enable-expert-parallel --all2all-backend deepep_low_latency --disable-uvicorn-access-log"
|
||||||
env:
|
env:
|
||||||
VLLM_USE_DEEP_GEMM: "1"
|
VLLM_USE_DEEP_GEMM: "1"
|
||||||
|
|||||||
+2
@@ -2,6 +2,8 @@ model_name: "RedHatAI/Qwen3-30B-A3B-FP8-block"
|
|||||||
accuracy_threshold: 0.85
|
accuracy_threshold: 0.85
|
||||||
num_questions: 1319
|
num_questions: 1319
|
||||||
num_fewshot: 5
|
num_fewshot: 5
|
||||||
|
# 30B MoE + DeepGEMM/DeepEP JIT compile can exceed the 1200s default startup wait.
|
||||||
|
startup_max_wait_seconds: 2400
|
||||||
server_args: "--enforce-eager --max-model-len 8192 --data-parallel-size 2 --enable-expert-parallel --all2all-backend deepep_low_latency --disable-uvicorn-access-log"
|
server_args: "--enforce-eager --max-model-len 8192 --data-parallel-size 2 --enable-expert-parallel --all2all-backend deepep_low_latency --disable-uvicorn-access-log"
|
||||||
env:
|
env:
|
||||||
VLLM_USE_DEEP_GEMM: "1"
|
VLLM_USE_DEEP_GEMM: "1"
|
||||||
|
|||||||
+2
@@ -2,4 +2,6 @@ model_name: "RedHatAI/Qwen3-30B-A3B-NVFP4"
|
|||||||
accuracy_threshold: 0.88
|
accuracy_threshold: 0.88
|
||||||
num_questions: 1319
|
num_questions: 1319
|
||||||
num_fewshot: 5
|
num_fewshot: 5
|
||||||
|
# 30B MoE + CuTeDSL/DeepEP JIT compile can exceed the 1200s default startup wait.
|
||||||
|
startup_max_wait_seconds: 2400
|
||||||
server_args: "--enforce-eager --max-model-len 8192 --data-parallel-size 2 --enable-expert-parallel --all2all-backend deepep_low_latency --moe-backend=flashinfer_cutedsl"
|
server_args: "--enforce-eager --max-model-len 8192 --data-parallel-size 2 --enable-expert-parallel --all2all-backend deepep_low_latency --moe-backend=flashinfer_cutedsl"
|
||||||
|
|||||||
+2
@@ -2,4 +2,6 @@ model_name: "nvidia/Qwen3-30B-A3B-NVFP4"
|
|||||||
accuracy_threshold: 0.88
|
accuracy_threshold: 0.88
|
||||||
num_questions: 1319
|
num_questions: 1319
|
||||||
num_fewshot: 5
|
num_fewshot: 5
|
||||||
|
# 30B MoE + CuTeDSL/DeepEP JIT compile can exceed the 1200s default startup wait.
|
||||||
|
startup_max_wait_seconds: 2400
|
||||||
server_args: "--enforce-eager --max-model-len 8192 --data-parallel-size 2 --enable-expert-parallel --all2all-backend deepep_low_latency --moe-backend=flashinfer_cutedsl"
|
server_args: "--enforce-eager --max-model-len 8192 --data-parallel-size 2 --enable-expert-parallel --all2all-backend deepep_low_latency --moe-backend=flashinfer_cutedsl"
|
||||||
|
|||||||
Reference in New Issue
Block a user