minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 10 Packages Projects Releases Wiki Activity

[AMD][CI] Update v4 CI setting and move the task to main branch (#26662)

Browse Source
This commit is contained in:
Thomas Wang
2026-05-29 00:12:40 -07:00
committed by GitHub
parent b2eed9e16d
commit 6e9bd82714
5 changed files with 109 additions and 172 deletions
Download Patch File Download Diff File Expand all files Collapse all files
.github/workflows/nightly-test-amd-rocm720.yml
+9 -64
View File
@@ -1162,52 +1162,20 @@ jobs:
with:
ref: ${{ inputs.ref || github.sha }}
- name: Resolve DSv4 image tag
id: dsv4_image
run: |
# Pick the latest Docker Hub tag matching rocm720-mi35x-<sha7>-<YYYYMMDD>-DSv4.
# Docker Hub returns results sorted by last_updated DESC by default, so the
# first regex match is the most recent daily build.
AUTH_HEADER=()
if [[ -n "${DOCKERHUB_AMD_USERNAME:-}" && -n "${DOCKERHUB_AMD_TOKEN:-}" ]]; then
TOKEN=$(curl -s -H "Content-Type: application/json" \
-X POST -d "{\"username\":\"${DOCKERHUB_AMD_USERNAME}\",\"password\":\"${DOCKERHUB_AMD_TOKEN}\"}" \
https://hub.docker.com/v2/users/login/ | python3 -c "import json,sys; print(json.load(sys.stdin).get('token',''))")
if [[ -n "$TOKEN" ]]; then
AUTH_HEADER=(-H "Authorization: JWT $TOKEN")
fi
fi
TAG=$(curl -s "${AUTH_HEADER[@]}" \
"https://hub.docker.com/v2/repositories/rocm/sgl-dev/tags?page_size=100&name=DSv4" \
| grep -oE '"name":"rocm720-mi35x-[a-f0-9]{7}-[0-9]{8}-DSv4"' \
| head -n 1 | cut -d'"' -f4)
if [ -z "$TAG" ]; then
echo "::error::No DSv4 image found matching rocm720-mi35x-<sha7>-<YYYYMMDD>-DSv4 on Docker Hub"
exit 1
fi
echo "image=rocm/sgl-dev:$TAG" >> "$GITHUB_OUTPUT"
echo "Resolved DSv4 image: rocm/sgl-dev:$TAG"
- name: Ensure VRAM is clear
run: bash scripts/ci/amd/ensure_vram_clear.sh rocm
- name: Setup docker (ROCm 7.2 DSv4)
- name: Setup docker (ROCm 7.2)
run: |
touch github_summary.md
bash scripts/ci/amd/amd_ci_start_container.sh --custom-image ${{ steps.dsv4_image.outputs.image }}
bash scripts/ci/amd/amd_ci_start_container.sh --rocm-version rocm720
env:
GITHUB_WORKSPACE: ${{ github.workspace }}
- name: Install dependencies (preserve DSv4 sglang/aiter from image)
- name: Install dependencies
run: |
# --skip-sglang-build: keep the image's pre-installed DSv4 sglang
# (default would `pip install -e /sglang-checkout/python` and clobber it with main's source).
# --skip-aiter-build: keep the image's DSv4-tuned aiter
# (default reads /sglang-checkout/docker/rocm.Dockerfile from main and rebuilds aiter to that commit).
# --skip-test-time-deps: GSM8K + bench_one_batch_server don't need lmms-eval / human-eval.
bash scripts/ci/amd/amd_ci_install_dependency.sh \
--skip-sglang-build --skip-aiter-build --skip-test-time-deps
# tabulate is the only thing run_suite.py imports that may not be in the DSv4 image.
bash scripts/ci/amd/amd_ci_install_dependency.sh --skip-test-time-deps
bash scripts/ci/amd/amd_ci_exec.sh pip install tabulate
- name: Accuracy + Performance Test MI35x ROCm 7.2 (8-GPU DeepSeek-V4-Flash FP8 + FP4)
@@ -1229,43 +1197,20 @@ jobs:
with:
ref: ${{ inputs.ref || github.sha }}
- name: Resolve DSv4 image tag
id: dsv4_image
run: |
AUTH_HEADER=()
if [[ -n "${DOCKERHUB_AMD_USERNAME:-}" && -n "${DOCKERHUB_AMD_TOKEN:-}" ]]; then
TOKEN=$(curl -s -H "Content-Type: application/json" \
-X POST -d "{\"username\":\"${DOCKERHUB_AMD_USERNAME}\",\"password\":\"${DOCKERHUB_AMD_TOKEN}\"}" \
https://hub.docker.com/v2/users/login/ | python3 -c "import json,sys; print(json.load(sys.stdin).get('token',''))")
if [[ -n "$TOKEN" ]]; then
AUTH_HEADER=(-H "Authorization: JWT $TOKEN")
fi
fi
TAG=$(curl -s "${AUTH_HEADER[@]}" \
"https://hub.docker.com/v2/repositories/rocm/sgl-dev/tags?page_size=100&name=DSv4" \
| grep -oE '"name":"rocm720-mi35x-[a-f0-9]{7}-[0-9]{8}-DSv4"' \
| head -n 1 | cut -d'"' -f4)
if [ -z "$TAG" ]; then
echo "::error::No DSv4 image found matching rocm720-mi35x-<sha7>-<YYYYMMDD>-DSv4 on Docker Hub"
exit 1
fi
echo "image=rocm/sgl-dev:$TAG" >> "$GITHUB_OUTPUT"
echo "Resolved DSv4 image: rocm/sgl-dev:$TAG"
- name: Ensure VRAM is clear
run: bash scripts/ci/amd/ensure_vram_clear.sh rocm
- name: Setup docker (ROCm 7.2 DSv4)
- name: Setup docker (ROCm 7.2)
run: |
touch github_summary.md
bash scripts/ci/amd/amd_ci_start_container.sh --custom-image ${{ steps.dsv4_image.outputs.image }}
bash scripts/ci/amd/amd_ci_start_container.sh --rocm-version rocm720
env:
GITHUB_WORKSPACE: ${{ github.workspace }}
- name: Install dependencies (preserve DSv4 sglang/aiter from image)
- name: Install dependencies
run: |
bash scripts/ci/amd/amd_ci_install_dependency.sh \
--skip-sglang-build --skip-aiter-build --skip-test-time-deps
# --skip-test-time-deps: GSM8K + bench_one_batch_server don't need lmms-eval / human-eval.
bash scripts/ci/amd/amd_ci_install_dependency.sh --skip-test-time-deps
bash scripts/ci/amd/amd_ci_exec.sh pip install tabulate
- name: Accuracy + Performance Test MI35x ROCm 7.2 (8-GPU DeepSeek-V4-Pro FP8 + FP4)
test/registered/amd/test_deepseek_v4_flash_fp4.py
+25 -27
View File
@@ -35,38 +35,32 @@ DEEPSEEK_V4_FP4_MODEL_PATH = os.environ.get(
)
SERVER_LAUNCH_TIMEOUT = 3600
# Common DeepSeek-V4 env vars (AMD ROCm 7.2 path: tilelang + AITER + ROCm700A).
# Source of truth: python/run_dsv4.sh.
# Common DeepSeek-V4 env vars (AMD ROCm 7.2 path: AITER indexer + triton attn + ROCm700A).
COMMON_ENV_VARS = {
"SGLANG_OPT_USE_FUSED_COMPRESS": "true",
"SGLANG_OPT_USE_OLD_COMPRESSOR": "true",
"SGLANG_OPT_USE_TILELANG_SWA_PREPARE": "false",
"SGLANG_OPT_USE_TRITON_SWA_PREPARE": "true",
"SGLANG_OPT_USE_JIT_KERNEL_FUSED_TOPK": "false",
"SGLANG_OPT_USE_FUSED_HASH_TOPK": "false",
"SGLANG_OPT_DEEPGEMM_HC_PRENORM": "false",
"SGLANG_OPT_USE_TILELANG_MHC_PRE": "false",
"SGLANG_OPT_USE_AITER_MHC_PRE": "true",
"SGLANG_OPT_USE_TILELANG_MHC_POST": "false",
"SGLANG_OPT_USE_AITER_MHC_POST": "true",
"SGLANG_ENABLE_THINKING": "1",
"SGLANG_USE_AITER": "1",
"AITER_BF16_FP8_MOE_BOUND": "1",
"SGLANG_USE_ROCM700A": "1",
"SGLANG_FP8_PAGED_MQA_LOGITS_TORCH": "1",
"SGLANG_OPT_DPSK_V4_RADIX": "0",
"SGLANG_OPT_USE_OVERLAP_STORE_CACHE": "false",
"SGLANG_OPT_USE_FUSED_STORE_CACHE": "false",
"SGLANG_TOPK_TRANSFORM_512_TORCH": "1",
"SGLANG_OPT_USE_TILELANG_INDEXER": "true",
"SGLANG_HACK_FLASHMLA_BACKEND": "triton",
"SGLANG_DEFAULT_THINKING": "1",
"SGLANG_DSV4_REASONING_EFFORT": "max",
"SGLANG_OPT_DEEPGEMM_HC_PRENORM": "false",
"SGLANG_USE_AITER": "1",
"SGLANG_USE_ROCM700A": "1",
"SGLANG_OPT_USE_FUSED_COMPRESS": "true",
"SGLANG_OPT_USE_FUSED_COMPRESS_TRITON": "true",
"SGLANG_HACK_FLASHMLA_BACKEND": "triton",
"SGLANG_OPT_FP8_WO_A_GEMM": "false",
"SGLANG_OPT_USE_JIT_INDEXER_METADATA": "false",
"SGLANG_OPT_USE_TOPK_V2": "false",
"SGLANG_OPT_USE_AITER_INDEXER": "true",
"SGLANG_OPT_USE_TILELANG_INDEXER": "false",
"SGLANG_OPT_USE_TILELANG_MHC_PRE": "false",
"SGLANG_OPT_USE_TILELANG_MHC_POST": "false",
"SGLANG_FP8_PAGED_MQA_LOGITS_TORCH": "1",
"SGLANG_OPT_USE_MULTI_STREAM_OVERLAP": "false",
"SGLANG_ROCM_USE_MULTI_STREAM": "false",
"AITER_BF16_FP8_MOE_BOUND": "0",
}
# FP4 variant: FP4 mixed-precision experts.
# FP4 variant
FP4_ENV_VARS = {
"SGLANG_DSV4_FP4_EXPERTS": "true",
"SGLANG_FORCE_TRITON_MOE_FP8": "0",
}
@@ -86,11 +80,15 @@ class TestDeepseekV4Fp4(CustomTestCase):
"8",
"--disable-radix-cache",
"--attention-backend",
"compressed",
"dsv4",
"--max-running-requests",
"256",
"--page-size",
"256",
"--mem-fraction-static",
"0.90",
"--swa-full-tokens-ratio",
"0.1",
"--chunked-prefill-size",
"8192",
"--disable-shared-experts-fusion",
test/registered/amd/test_deepseek_v4_flash_fp8.py
+25 -27
View File
@@ -35,38 +35,32 @@ DEEPSEEK_V4_FP8_MODEL_PATH = os.environ.get(
)
SERVER_LAUNCH_TIMEOUT = 3600
# Common DeepSeek-V4 env vars (AMD ROCm 7.2 path: tilelang + AITER + ROCm700A).
# Source of truth: python/run_dsv4.sh.
# Common DeepSeek-V4 env vars (AMD ROCm 7.2 path: AITER indexer + triton attn + ROCm700A).
COMMON_ENV_VARS = {
"SGLANG_OPT_USE_FUSED_COMPRESS": "true",
"SGLANG_OPT_USE_OLD_COMPRESSOR": "true",
"SGLANG_OPT_USE_TILELANG_SWA_PREPARE": "false",
"SGLANG_OPT_USE_TRITON_SWA_PREPARE": "true",
"SGLANG_OPT_USE_JIT_KERNEL_FUSED_TOPK": "false",
"SGLANG_OPT_USE_FUSED_HASH_TOPK": "false",
"SGLANG_OPT_DEEPGEMM_HC_PRENORM": "false",
"SGLANG_OPT_USE_TILELANG_MHC_PRE": "false",
"SGLANG_OPT_USE_AITER_MHC_PRE": "true",
"SGLANG_OPT_USE_TILELANG_MHC_POST": "false",
"SGLANG_OPT_USE_AITER_MHC_POST": "true",
"SGLANG_ENABLE_THINKING": "1",
"SGLANG_USE_AITER": "1",
"AITER_BF16_FP8_MOE_BOUND": "1",
"SGLANG_USE_ROCM700A": "1",
"SGLANG_FP8_PAGED_MQA_LOGITS_TORCH": "1",
"SGLANG_OPT_DPSK_V4_RADIX": "0",
"SGLANG_OPT_USE_OVERLAP_STORE_CACHE": "false",
"SGLANG_OPT_USE_FUSED_STORE_CACHE": "false",
"SGLANG_TOPK_TRANSFORM_512_TORCH": "1",
"SGLANG_OPT_USE_TILELANG_INDEXER": "true",
"SGLANG_HACK_FLASHMLA_BACKEND": "triton",
"SGLANG_DEFAULT_THINKING": "1",
"SGLANG_DSV4_REASONING_EFFORT": "max",
"SGLANG_OPT_DEEPGEMM_HC_PRENORM": "false",
"SGLANG_USE_AITER": "1",
"SGLANG_USE_ROCM700A": "1",
"SGLANG_OPT_USE_FUSED_COMPRESS": "true",
"SGLANG_OPT_USE_FUSED_COMPRESS_TRITON": "true",
"SGLANG_HACK_FLASHMLA_BACKEND": "triton",
"SGLANG_OPT_FP8_WO_A_GEMM": "false",
"SGLANG_OPT_USE_JIT_INDEXER_METADATA": "false",
"SGLANG_OPT_USE_TOPK_V2": "false",
"SGLANG_OPT_USE_AITER_INDEXER": "true",
"SGLANG_OPT_USE_TILELANG_INDEXER": "false",
"SGLANG_OPT_USE_TILELANG_MHC_PRE": "false",
"SGLANG_OPT_USE_TILELANG_MHC_POST": "false",
"SGLANG_FP8_PAGED_MQA_LOGITS_TORCH": "1",
"SGLANG_OPT_USE_MULTI_STREAM_OVERLAP": "false",
"SGLANG_ROCM_USE_MULTI_STREAM": "false",
"AITER_BF16_FP8_MOE_BOUND": "0",
}
# FP8 variant: dense-FP8 experts via the Triton MoE FP8 path.
# FP8 variant
FP8_ENV_VARS = {
"SGLANG_DSV4_FP4_EXPERTS": "false",
"SGLANG_FORCE_TRITON_MOE_FP8": "1",
}
@@ -86,11 +80,15 @@ class TestDeepseekV4Fp8(CustomTestCase):
"8",
"--disable-radix-cache",
"--attention-backend",
"compressed",
"dsv4",
"--max-running-requests",
"256",
"--page-size",
"256",
"--mem-fraction-static",
"0.90",
"--swa-full-tokens-ratio",
"0.1",
"--chunked-prefill-size",
"8192",
"--disable-shared-experts-fusion",
test/registered/amd/test_deepseek_v4_pro_fp4.py
+25 -27
View File
@@ -37,38 +37,32 @@ DEEPSEEK_V4_PRO_FP4_MODEL_PATH = os.environ.get(
# Pro is 1.6T; weight load + warmup is much longer than Flash 285B.
SERVER_LAUNCH_TIMEOUT = 5400
# Common DeepSeek-V4 env vars (AMD ROCm 7.2 path: tilelang + AITER + ROCm700A).
# Source of truth: python/run_dsv4.sh.
# Common DeepSeek-V4 env vars (AMD ROCm 7.2 path: AITER indexer + triton attn + ROCm700A).
COMMON_ENV_VARS = {
"SGLANG_OPT_USE_FUSED_COMPRESS": "true",
"SGLANG_OPT_USE_OLD_COMPRESSOR": "true",
"SGLANG_OPT_USE_TILELANG_SWA_PREPARE": "false",
"SGLANG_OPT_USE_TRITON_SWA_PREPARE": "true",
"SGLANG_OPT_USE_JIT_KERNEL_FUSED_TOPK": "false",
"SGLANG_OPT_USE_FUSED_HASH_TOPK": "false",
"SGLANG_OPT_DEEPGEMM_HC_PRENORM": "false",
"SGLANG_OPT_USE_TILELANG_MHC_PRE": "false",
"SGLANG_OPT_USE_AITER_MHC_PRE": "true",
"SGLANG_OPT_USE_TILELANG_MHC_POST": "false",
"SGLANG_OPT_USE_AITER_MHC_POST": "true",
"SGLANG_ENABLE_THINKING": "1",
"SGLANG_USE_AITER": "1",
"AITER_BF16_FP8_MOE_BOUND": "1",
"SGLANG_USE_ROCM700A": "1",
"SGLANG_FP8_PAGED_MQA_LOGITS_TORCH": "1",
"SGLANG_OPT_DPSK_V4_RADIX": "0",
"SGLANG_OPT_USE_OVERLAP_STORE_CACHE": "false",
"SGLANG_OPT_USE_FUSED_STORE_CACHE": "false",
"SGLANG_TOPK_TRANSFORM_512_TORCH": "1",
"SGLANG_OPT_USE_TILELANG_INDEXER": "true",
"SGLANG_HACK_FLASHMLA_BACKEND": "triton",
"SGLANG_DEFAULT_THINKING": "1",
"SGLANG_DSV4_REASONING_EFFORT": "max",
"SGLANG_OPT_DEEPGEMM_HC_PRENORM": "false",
"SGLANG_USE_AITER": "1",
"SGLANG_USE_ROCM700A": "1",
"SGLANG_OPT_USE_FUSED_COMPRESS": "true",
"SGLANG_OPT_USE_FUSED_COMPRESS_TRITON": "true",
"SGLANG_HACK_FLASHMLA_BACKEND": "triton",
"SGLANG_OPT_FP8_WO_A_GEMM": "false",
"SGLANG_OPT_USE_JIT_INDEXER_METADATA": "false",
"SGLANG_OPT_USE_TOPK_V2": "false",
"SGLANG_OPT_USE_AITER_INDEXER": "true",
"SGLANG_OPT_USE_TILELANG_INDEXER": "false",
"SGLANG_OPT_USE_TILELANG_MHC_PRE": "false",
"SGLANG_OPT_USE_TILELANG_MHC_POST": "false",
"SGLANG_FP8_PAGED_MQA_LOGITS_TORCH": "1",
"SGLANG_OPT_USE_MULTI_STREAM_OVERLAP": "false",
"SGLANG_ROCM_USE_MULTI_STREAM": "false",
"AITER_BF16_FP8_MOE_BOUND": "0",
}
# FP4 variant: FP4 mixed-precision experts.
# FP4 variant
FP4_ENV_VARS = {
"SGLANG_DSV4_FP4_EXPERTS": "true",
"SGLANG_FORCE_TRITON_MOE_FP8": "0",
}
@@ -88,11 +82,15 @@ class TestDeepseekV4ProFp4(CustomTestCase):
"8",
"--disable-radix-cache",
"--attention-backend",
"compressed",
"dsv4",
"--max-running-requests",
"256",
"--page-size",
"256",
"--mem-fraction-static",
"0.90",
"--swa-full-tokens-ratio",
"0.1",
"--chunked-prefill-size",
"8192",
"--disable-shared-experts-fusion",
test/registered/amd/test_deepseek_v4_pro_fp8.py
+25 -27
View File
@@ -37,38 +37,32 @@ DEEPSEEK_V4_PRO_FP8_MODEL_PATH = os.environ.get(
# Pro is 1.6T; weight load + warmup is much longer than Flash 285B.
SERVER_LAUNCH_TIMEOUT = 5400
# Common DeepSeek-V4 env vars (AMD ROCm 7.2 path: tilelang + AITER + ROCm700A).
# Source of truth: python/run_dsv4.sh.
# Common DeepSeek-V4 env vars (AMD ROCm 7.2 path: AITER indexer + triton attn + ROCm700A).
COMMON_ENV_VARS = {
"SGLANG_OPT_USE_FUSED_COMPRESS": "true",
"SGLANG_OPT_USE_OLD_COMPRESSOR": "true",
"SGLANG_OPT_USE_TILELANG_SWA_PREPARE": "false",
"SGLANG_OPT_USE_TRITON_SWA_PREPARE": "true",
"SGLANG_OPT_USE_JIT_KERNEL_FUSED_TOPK": "false",
"SGLANG_OPT_USE_FUSED_HASH_TOPK": "false",
"SGLANG_OPT_DEEPGEMM_HC_PRENORM": "false",
"SGLANG_OPT_USE_TILELANG_MHC_PRE": "false",
"SGLANG_OPT_USE_AITER_MHC_PRE": "true",
"SGLANG_OPT_USE_TILELANG_MHC_POST": "false",
"SGLANG_OPT_USE_AITER_MHC_POST": "true",
"SGLANG_ENABLE_THINKING": "1",
"SGLANG_USE_AITER": "1",
"AITER_BF16_FP8_MOE_BOUND": "1",
"SGLANG_USE_ROCM700A": "1",
"SGLANG_FP8_PAGED_MQA_LOGITS_TORCH": "1",
"SGLANG_OPT_DPSK_V4_RADIX": "0",
"SGLANG_OPT_USE_OVERLAP_STORE_CACHE": "false",
"SGLANG_OPT_USE_FUSED_STORE_CACHE": "false",
"SGLANG_TOPK_TRANSFORM_512_TORCH": "1",
"SGLANG_OPT_USE_TILELANG_INDEXER": "true",
"SGLANG_HACK_FLASHMLA_BACKEND": "triton",
"SGLANG_DEFAULT_THINKING": "1",
"SGLANG_DSV4_REASONING_EFFORT": "max",
"SGLANG_OPT_DEEPGEMM_HC_PRENORM": "false",
"SGLANG_USE_AITER": "1",
"SGLANG_USE_ROCM700A": "1",
"SGLANG_OPT_USE_FUSED_COMPRESS": "true",
"SGLANG_OPT_USE_FUSED_COMPRESS_TRITON": "true",
"SGLANG_HACK_FLASHMLA_BACKEND": "triton",
"SGLANG_OPT_FP8_WO_A_GEMM": "false",
"SGLANG_OPT_USE_JIT_INDEXER_METADATA": "false",
"SGLANG_OPT_USE_TOPK_V2": "false",
"SGLANG_OPT_USE_AITER_INDEXER": "true",
"SGLANG_OPT_USE_TILELANG_INDEXER": "false",
"SGLANG_OPT_USE_TILELANG_MHC_PRE": "false",
"SGLANG_OPT_USE_TILELANG_MHC_POST": "false",
"SGLANG_FP8_PAGED_MQA_LOGITS_TORCH": "1",
"SGLANG_OPT_USE_MULTI_STREAM_OVERLAP": "false",
"SGLANG_ROCM_USE_MULTI_STREAM": "false",
"AITER_BF16_FP8_MOE_BOUND": "0",
}
# FP8 variant: dense-FP8 experts via the Triton MoE FP8 path.
# FP8 variant
FP8_ENV_VARS = {
"SGLANG_DSV4_FP4_EXPERTS": "false",
"SGLANG_FORCE_TRITON_MOE_FP8": "1",
}
@@ -88,11 +82,15 @@ class TestDeepseekV4ProFp8(CustomTestCase):
"8",
"--disable-radix-cache",
"--attention-backend",
"compressed",
"dsv4",
"--max-running-requests",
"256",
"--page-size",
"256",
"--mem-fraction-static",
"0.90",
"--swa-full-tokens-ratio",
"0.1",
"--chunked-prefill-size",
"8192",
"--disable-shared-experts-fusion",