Skip to content

Commit 8602020

Browse files
committed
revert(evals): preserve benchmark recipes
1 parent 9d4976b commit 8602020

2 files changed

Lines changed: 7 additions & 12 deletions

File tree

benchmarks/single_node/agentic/dsv4_fp4_mi355x_sglang.sh

Lines changed: 2 additions & 7 deletions
Original file line numberDiff line numberDiff line change
@@ -110,13 +110,8 @@ MAX_RUNNING_REQUESTS=$((2 * CONC))
110110
CUDA_GRAPH_MAX_BS=$CONC
111111
[ "$CUDA_GRAPH_MAX_BS" -gt 128 ] && CUDA_GRAPH_MAX_BS=128
112112

113-
if [ "${IS_AGENTIC:-0}" = "1" ]; then
114-
export SGLANG_DEFAULT_THINKING=1
115-
export SGLANG_DSV4_REASONING_EFFORT=high
116-
else
117-
export SGLANG_DEFAULT_THINKING=1
118-
export SGLANG_DSV4_REASONING_EFFORT=max
119-
fi
113+
export SGLANG_DEFAULT_THINKING=1
114+
export SGLANG_DSV4_REASONING_EFFORT=high
120115
export SGLANG_USE_ROCM700A=0
121116
export SGLANG_HACK_FLASHMLA_BACKEND=unified_kv_triton
122117
export AITER_BF16_FP8_MOE_BOUND=0

configs/amd-master.yaml

Lines changed: 5 additions & 5 deletions
Original file line numberDiff line numberDiff line change
@@ -1915,7 +1915,7 @@ dsv4-fp4-mi355x-sglang-disagg:
19151915

19161916

19171917
dsv4-fp4-mi355x-sglang:
1918-
image: lmsysorg/sglang-rocm:v0.5.14-rocm720-mi35x-20260708
1918+
image: lmsysorg/sglang-rocm:v0.5.14-rocm720-mi35x-20260706
19191919
model: deepseek-ai/DeepSeek-V4-Pro
19201920
model-prefix: dsv4
19211921
runner: mi355x
@@ -1984,7 +1984,7 @@ dsv4-fp4-mi355x-sglang-mtp:
19841984
# gpu-mem-util=0.6. TP8 sweeps conc 4-64; DEP8 has a single conc=64
19851985
# probe to validate the ROCm DP+EP path.
19861986
dsv4-fp4-mi355x-vllm:
1987-
image: vllm/vllm-openai-rocm:nightly-67fe73b2b47e2927af1e5ec6fb91e16e3b075940
1987+
image: vllm/vllm-openai-rocm:nightly-09663abde0f50944a8d5ea30120666024b503faa
19881988
model: deepseek-ai/DeepSeek-V4-Pro
19891989
model-prefix: dsv4
19901990
runner: mi355x
@@ -2008,10 +2008,10 @@ dsv4-fp4-mi355x-vllm:
20082008
# vllm-project/vllm#43385 (ROCm DeepSeek-V4 MTP, merged 2026-05-24, included in
20092009
# v0.22.0). Full conc 4-512 range maps the complete crossover curve: MTP wins
20102010
# at low batch (PR perf data: +75% @ conc1, +38% @ conc8) and falls behind STP
2011-
# above ~conc32 (-37% @ conc32). The pinned nightly contains both the base
2012-
# DeepSeek-V4 ROCm path and the MTP implementation.
2011+
# above ~conc32 (-37% @ conc32). Image reuses the base entry's v0.22.0 ROCm
2012+
# build, which already contains the MTP commit.
20132013
dsv4-fp4-mi355x-vllm-mtp:
2014-
image: vllm/vllm-openai-rocm:nightly-67fe73b2b47e2927af1e5ec6fb91e16e3b075940
2014+
image: vllm/vllm-openai-rocm:nightly-09663abde0f50944a8d5ea30120666024b503faa
20152015
model: deepseek-ai/DeepSeek-V4-Pro
20162016
model-prefix: dsv4
20172017
runner: mi355x

0 commit comments

Comments
 (0)