From d090c56b8b92e92dcff114835cf922983ed105d8 Mon Sep 17 00:00:00 2001 From: Cam Quilici Date: Fri, 17 Apr 2026 11:27:36 -0500 Subject: [PATCH] Revert "[AMD][MI30X]Update Qwen3.5 perf (#986)" [skip-sweep] This reverts commit 20073bafdaae43b116a50ec6985e01b945755a7f, except for changes to benchmarks/single_node/qwen3.5_{bf16,fp8}_mi355x.sh, which have been preserved to retain PR #1036's subsequent fixes. --- .github/configs/amd-master.yaml | 8 ++++---- benchmarks/single_node/qwen3.5_bf16_mi300x.sh | 15 +++------------ benchmarks/single_node/qwen3.5_bf16_mi325x.sh | 15 +++------------ benchmarks/single_node/qwen3.5_fp8_mi300x.sh | 15 +++------------ benchmarks/single_node/qwen3.5_fp8_mi325x.sh | 15 +++------------ perf-changelog.yaml | 10 ---------- 6 files changed, 16 insertions(+), 62 deletions(-) diff --git a/.github/configs/amd-master.yaml b/.github/configs/amd-master.yaml index b35716e78..1f1093e38 100644 --- a/.github/configs/amd-master.yaml +++ b/.github/configs/amd-master.yaml @@ -132,7 +132,7 @@ qwen3.5-bf16-mi355x-sglang: - { tp: 8, ep: 1, conc-start: 4, conc-end: 256 } qwen3.5-bf16-mi300x-sglang: - image: lmsysorg/sglang:v0.5.10-rocm720-mi30x + image: lmsysorg/sglang:v0.5.9-rocm720-mi30x model: Qwen/Qwen3.5-397B-A17B model-prefix: qwen3.5 runner: mi300x @@ -150,7 +150,7 @@ qwen3.5-bf16-mi300x-sglang: - { tp: 8, conc-start: 4, conc-end: 64 } qwen3.5-bf16-mi325x-sglang: - image: lmsysorg/sglang:v0.5.10-rocm720-mi30x + image: lmsysorg/sglang:v0.5.9-rocm720-mi30x model: Qwen/Qwen3.5-397B-A17B model-prefix: qwen3.5 runner: mi325x @@ -168,7 +168,7 @@ qwen3.5-bf16-mi325x-sglang: - { tp: 8, conc-start: 4, conc-end: 64 } qwen3.5-fp8-mi325x-sglang: - image: lmsysorg/sglang:v0.5.10-rocm720-mi30x + image: lmsysorg/sglang:v0.5.9-rocm720-mi30x model: Qwen/Qwen3.5-397B-A17B-FP8 model-prefix: qwen3.5 runner: mi325x @@ -227,7 +227,7 @@ qwen3.5-fp4-mi355x-sglang: - { tp: 4, conc-start: 4, conc-end: 16 } qwen3.5-fp8-mi300x-sglang: - image: lmsysorg/sglang:v0.5.10-rocm720-mi30x + image: lmsysorg/sglang:v0.5.9-rocm720-mi30x model: Qwen/Qwen3.5-397B-A17B-FP8 model-prefix: qwen3.5 runner: mi300x diff --git a/benchmarks/single_node/qwen3.5_bf16_mi300x.sh b/benchmarks/single_node/qwen3.5_bf16_mi300x.sh index f7c71963d..8aca9860a 100755 --- a/benchmarks/single_node/qwen3.5_bf16_mi300x.sh +++ b/benchmarks/single_node/qwen3.5_bf16_mi300x.sh @@ -19,34 +19,25 @@ hf download "$MODEL" SERVER_LOG=/workspace/server.log PORT=${PORT:-8888} -CONTEXT_LENGTH=$((ISL + OSL + 20)) -MAX_PREFILL_TOKENS=32768 EVAL_CONTEXT_ARGS="" if [ "${EVAL_ONLY}" = "true" ]; then setup_eval_context EVAL_CONTEXT_ARGS="--context-length $EVAL_MAX_MODEL_LEN" -else EVAL_CONTEXT_ARGS="--context-length $CONTEXT_LENGTH" fi # Start GPU monitoring (power, temperature, clocks every second) start_gpu_monitor # following Andy Luo linkedin's recipe https://www.linkedin.com/feed/update/urn:li:activity:7429203734389280768/ python3 -m sglang.launch_server \ - --attention-backend aiter \ + --attention-backend triton \ --model-path $MODEL \ --host=0.0.0.0 \ --port $PORT \ --tensor-parallel-size $TP \ - --data-parallel-size 1 \ --trust-remote-code \ - --tokenizer-worker-num 6 \ - --enable-aiter-allreduce-fusion \ - --cuda-graph-max-bs $CONC \ - --disable-radix-cache \ - --max-prefill-tokens $MAX_PREFILL_TOKENS \ - --scheduler-recv-interval 30 \ - --mem-fraction-static 0.75 $EVAL_CONTEXT_ARGS > $SERVER_LOG 2>&1 & + --mem-fraction-static 0.8 \ + --disable-radix-cache $EVAL_CONTEXT_ARGS > $SERVER_LOG 2>&1 & SERVER_PID=$! diff --git a/benchmarks/single_node/qwen3.5_bf16_mi325x.sh b/benchmarks/single_node/qwen3.5_bf16_mi325x.sh index f7c71963d..8aca9860a 100644 --- a/benchmarks/single_node/qwen3.5_bf16_mi325x.sh +++ b/benchmarks/single_node/qwen3.5_bf16_mi325x.sh @@ -19,34 +19,25 @@ hf download "$MODEL" SERVER_LOG=/workspace/server.log PORT=${PORT:-8888} -CONTEXT_LENGTH=$((ISL + OSL + 20)) -MAX_PREFILL_TOKENS=32768 EVAL_CONTEXT_ARGS="" if [ "${EVAL_ONLY}" = "true" ]; then setup_eval_context EVAL_CONTEXT_ARGS="--context-length $EVAL_MAX_MODEL_LEN" -else EVAL_CONTEXT_ARGS="--context-length $CONTEXT_LENGTH" fi # Start GPU monitoring (power, temperature, clocks every second) start_gpu_monitor # following Andy Luo linkedin's recipe https://www.linkedin.com/feed/update/urn:li:activity:7429203734389280768/ python3 -m sglang.launch_server \ - --attention-backend aiter \ + --attention-backend triton \ --model-path $MODEL \ --host=0.0.0.0 \ --port $PORT \ --tensor-parallel-size $TP \ - --data-parallel-size 1 \ --trust-remote-code \ - --tokenizer-worker-num 6 \ - --enable-aiter-allreduce-fusion \ - --cuda-graph-max-bs $CONC \ - --disable-radix-cache \ - --max-prefill-tokens $MAX_PREFILL_TOKENS \ - --scheduler-recv-interval 30 \ - --mem-fraction-static 0.75 $EVAL_CONTEXT_ARGS > $SERVER_LOG 2>&1 & + --mem-fraction-static 0.8 \ + --disable-radix-cache $EVAL_CONTEXT_ARGS > $SERVER_LOG 2>&1 & SERVER_PID=$! diff --git a/benchmarks/single_node/qwen3.5_fp8_mi300x.sh b/benchmarks/single_node/qwen3.5_fp8_mi300x.sh index fe761d88d..00cc9cf91 100755 --- a/benchmarks/single_node/qwen3.5_fp8_mi300x.sh +++ b/benchmarks/single_node/qwen3.5_fp8_mi300x.sh @@ -19,14 +19,11 @@ hf download "$MODEL" SERVER_LOG=/workspace/server.log PORT=${PORT:-8888} -CONTEXT_LENGTH=$((ISL + OSL + 20)) -MAX_PREFILL_TOKENS=32768 EVAL_CONTEXT_ARGS="" if [ "${EVAL_ONLY}" = "true" ]; then setup_eval_context EVAL_CONTEXT_ARGS="--context-length $EVAL_MAX_MODEL_LEN" -else EVAL_CONTEXT_ARGS="--context-length $CONTEXT_LENGTH" fi # Start GPU monitoring (power, temperature, clocks every second) start_gpu_monitor @@ -34,20 +31,14 @@ start_gpu_monitor # following AMD Andy linkedin's recipe # https://www.linkedin.com/feed/update/urn:li:activity:7429203734389280768/ python3 -m sglang.launch_server \ - --attention-backend aiter \ + --attention-backend triton \ --model-path $MODEL \ --host=0.0.0.0 \ --port $PORT \ --tensor-parallel-size $TP \ - --data-parallel-size 1 \ --trust-remote-code \ - --tokenizer-worker-num 6 \ - --enable-aiter-allreduce-fusion \ - --cuda-graph-max-bs $CONC \ - --disable-radix-cache \ - --max-prefill-tokens $MAX_PREFILL_TOKENS \ - --scheduler-recv-interval 30 \ - --mem-fraction-static 0.75 $EVAL_CONTEXT_ARGS > $SERVER_LOG 2>&1 & + --mem-fraction-static 0.8 \ + --disable-radix-cache $EVAL_CONTEXT_ARGS > $SERVER_LOG 2>&1 & SERVER_PID=$! diff --git a/benchmarks/single_node/qwen3.5_fp8_mi325x.sh b/benchmarks/single_node/qwen3.5_fp8_mi325x.sh index fe761d88d..00cc9cf91 100755 --- a/benchmarks/single_node/qwen3.5_fp8_mi325x.sh +++ b/benchmarks/single_node/qwen3.5_fp8_mi325x.sh @@ -19,14 +19,11 @@ hf download "$MODEL" SERVER_LOG=/workspace/server.log PORT=${PORT:-8888} -CONTEXT_LENGTH=$((ISL + OSL + 20)) -MAX_PREFILL_TOKENS=32768 EVAL_CONTEXT_ARGS="" if [ "${EVAL_ONLY}" = "true" ]; then setup_eval_context EVAL_CONTEXT_ARGS="--context-length $EVAL_MAX_MODEL_LEN" -else EVAL_CONTEXT_ARGS="--context-length $CONTEXT_LENGTH" fi # Start GPU monitoring (power, temperature, clocks every second) start_gpu_monitor @@ -34,20 +31,14 @@ start_gpu_monitor # following AMD Andy linkedin's recipe # https://www.linkedin.com/feed/update/urn:li:activity:7429203734389280768/ python3 -m sglang.launch_server \ - --attention-backend aiter \ + --attention-backend triton \ --model-path $MODEL \ --host=0.0.0.0 \ --port $PORT \ --tensor-parallel-size $TP \ - --data-parallel-size 1 \ --trust-remote-code \ - --tokenizer-worker-num 6 \ - --enable-aiter-allreduce-fusion \ - --cuda-graph-max-bs $CONC \ - --disable-radix-cache \ - --max-prefill-tokens $MAX_PREFILL_TOKENS \ - --scheduler-recv-interval 30 \ - --mem-fraction-static 0.75 $EVAL_CONTEXT_ARGS > $SERVER_LOG 2>&1 & + --mem-fraction-static 0.8 \ + --disable-radix-cache $EVAL_CONTEXT_ARGS > $SERVER_LOG 2>&1 & SERVER_PID=$! diff --git a/perf-changelog.yaml b/perf-changelog.yaml index bc58bde07..f4ba37423 100644 --- a/perf-changelog.yaml +++ b/perf-changelog.yaml @@ -1314,16 +1314,6 @@ - "Configs: 1k1k (TP4 conc 4-128), 8k1k (TP4 conc 4-128)" pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/820 -- config-keys: - - qwen3.5-bf16-mi300x-sglang - - qwen3.5-bf16-mi325x-sglang - - qwen3.5-fp8-mi300x-sglang - - qwen3.5-fp8-mi325x-sglang - description: - - "Update cli args of Qwen3.5 FP8 and BF16 SGLang benchmarks for MI300X and MI325X to achieve better performance" - - "Use lmsysorg/sglang:v0.5.10-rocm720-mi30x" - pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/986 - - config-keys: - glm5-fp4-b200-sglang description: