From 0d441235f05804cb48c237da9617df66b5e22c7f Mon Sep 17 00:00:00 2001 From: ppalanga Date: Mon, 15 Dec 2025 06:48:58 -0800 Subject: [PATCH 1/4] Update amd-master.yaml --- .github/configs/amd-master.yaml | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/.github/configs/amd-master.yaml b/.github/configs/amd-master.yaml index a2674153a..a61f4fdd3 100644 --- a/.github/configs/amd-master.yaml +++ b/.github/configs/amd-master.yaml @@ -1,5 +1,5 @@ dsr1-fp4-mi355x-sglang: - image: rocm/7.0:rocm7.0_ubuntu_22.04_sgl-dev-v0.5.2-rocm7.0-mi35x-20250915 + image: lmsysorg/sglang:v0.5.6.post1-rocm700-mi35x model: amd/DeepSeek-R1-0528-MXFP4-Preview model-prefix: dsr1 runner: mi355x From 44aec220cca01c2c1210cfa9cb10e76e509bd5fa Mon Sep 17 00:00:00 2001 From: ppalanga Date: Mon, 15 Dec 2025 08:20:26 -0800 Subject: [PATCH 2/4] Update perf-changelog.yaml --- perf-changelog.yaml | 7 ++++++- 1 file changed, 6 insertions(+), 1 deletion(-) diff --git a/perf-changelog.yaml b/perf-changelog.yaml index a74285c53..926ac7e1d 100644 --- a/perf-changelog.yaml +++ b/perf-changelog.yaml @@ -80,4 +80,9 @@ description: | - Update vLLM image for NVIDIA configs from vLLM 0.11.0 to vLLM 0.11.2 - Adds kv-cache-dtype: fp8 to benchmarks/gptoss_fp4_b200_docker.sh - PR: https://github.com/InferenceMAX/InferenceMAX/pull/273 \ No newline at end of file + PR: https://github.com/InferenceMAX/InferenceMAX/pull/273 +- config-keys: + - dsr1-fp4-mi355x-sglang + description: | + - Updating MI355x Deepseek-R1 FP4 SGLang Image to upstream v0.5.6.post1 + PR: https://github.com/InferenceMAX/InferenceMAX/pull/330 From 1f9896506e2a9943aaddb3ef84f75cacd1328706 Mon Sep 17 00:00:00 2001 From: ppalanga Date: Mon, 15 Dec 2025 08:37:12 -0800 Subject: [PATCH 3/4] Update dsr1_fp4_mi355x_docker.sh --- benchmarks/dsr1_fp4_mi355x_docker.sh | 4 +++- 1 file changed, 3 insertions(+), 1 deletion(-) diff --git a/benchmarks/dsr1_fp4_mi355x_docker.sh b/benchmarks/dsr1_fp4_mi355x_docker.sh index ca1255802..14302715a 100644 --- a/benchmarks/dsr1_fp4_mi355x_docker.sh +++ b/benchmarks/dsr1_fp4_mi355x_docker.sh @@ -30,7 +30,9 @@ python3 -m sglang.launch_server --model-path=$MODEL --trust-remote-code \ --disable-radix-cache \ --num-continuous-decode-steps=4 \ --max-prefill-tokens=$PREFILL_SIZE \ ---cuda-graph-max-bs=128 > $SERVER_LOG 2>&1 & +--cuda-graph-max-bs=128 \ +--attention-backend aiter \ +--kv-cache-dtype fp8_e4m3 > $SERVER_LOG 2>&1 & SERVER_PID=$! From bf5ea3d5f6a178bd80b492359233b98514153723 Mon Sep 17 00:00:00 2001 From: ppalanga Date: Mon, 15 Dec 2025 13:29:57 -0800 Subject: [PATCH 4/4] Update dsr1_fp4_mi355x_docker.sh --- benchmarks/dsr1_fp4_mi355x_docker.sh | 1 + 1 file changed, 1 insertion(+) diff --git a/benchmarks/dsr1_fp4_mi355x_docker.sh b/benchmarks/dsr1_fp4_mi355x_docker.sh index 14302715a..8b3750a64 100644 --- a/benchmarks/dsr1_fp4_mi355x_docker.sh +++ b/benchmarks/dsr1_fp4_mi355x_docker.sh @@ -11,6 +11,7 @@ # RESULT_FILENAME # NUM_PROMPTS export SGLANG_USE_AITER=1 +export ROCM_QUICK_REDUCE_QUANTIZATION=INT4 PREFILL_SIZE=196608 if [[ "$ISL" == "8192" && "$OSL" == "1024" ]]; then