Skip to content

Commit efac275

Browse files
seungrokjclaude
andcommitted
[AMD] [AGENTX] dsv4-fp4-mi355x-sglang-mtp: gate simulated acceptance to throughput runs
Only export SGLANG_SIMULATE_ACC_* when EVAL_ONLY=false. Simulated acceptance bypasses real target verification, so it must be off for accuracy (EVAL_ONLY) runs or it corrupts the eval score. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
1 parent 782c075 commit efac275

1 file changed

Lines changed: 6 additions & 3 deletions

File tree

benchmarks/single_node/agentic/dsv4_fp4_mi355x_sglang_mtp.sh

Lines changed: 6 additions & 3 deletions
Original file line numberDiff line numberDiff line change
@@ -118,9 +118,12 @@ CUDA_GRAPH_MAX_BS=$CONC
118118
# openai.BadRequestError: Error code: 400 - {'object': 'error', 'message': "The input (3620936 tokens) is longer than the model's context length (1048576 tokens).", 'type': 'BadRequestError', 'param': None, 'code': 400}
119119
export SGLANG_DEFAULT_THINKING=1
120120
export SGLANG_DSV4_REASONING_EFFORT=high
121-
export SGLANG_SIMULATE_ACC_LEN=2.49
122-
export SGLANG_SIMULATE_ACC_METHOD=match-expected
123-
export SGLANG_SIMULATE_ACC_TOKEN_MODE=real-draft-token
121+
122+
if [ "${EVAL_ONLY:-false}" = "false" ]; then
123+
export SGLANG_SIMULATE_ACC_LEN=2.49
124+
export SGLANG_SIMULATE_ACC_METHOD=match-expected
125+
export SGLANG_SIMULATE_ACC_TOKEN_MODE=real-draft-token
126+
fi
124127

125128
export SGLANG_USE_ROCM700A=0
126129
export SGLANG_HACK_FLASHMLA_BACKEND=unified_kv_triton

0 commit comments

Comments
 (0)