Add tests to AMD CI for MI35x (#9662)

Co-authored-by: Sai Enduri <saimanas.enduri@amd.com>

.github/workflows/pr-test-amd.yml

@@ -28,6 +28,7 @@ jobs:
     if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') &&
       github.event.pull_request.draft == false
     strategy:
+      fail-fast: false
       matrix:
         runner: [linux-mi300-gpu-1, linux-mi325-gpu-1]
     runs-on: ${{matrix.runner}}
@@ -54,8 +55,9 @@ jobs:
     if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') &&
         github.event.pull_request.draft == false
     strategy:
+      fail-fast: false
       matrix:
-        runner: [linux-mi300-gpu-2, linux-mi325-gpu-2]
+        runner: [linux-mi300-gpu-2, linux-mi325-gpu-2, linux-mi35x-gpu-2]
     runs-on: ${{matrix.runner}}
     steps:
       - name: Checkout code
@@ -70,7 +72,7 @@ jobs:
         run: bash scripts/ci/amd_ci_install_dependency.sh
 
       - name: Evaluate accuracy (TP=2)
-        timeout-minutes: 30
+        timeout-minutes: 60
         run: |
           bash scripts/ci/amd_ci_exec.sh -e SGLANG_USE_AITER=0 python3 test_moe_eval_accuracy_large.py
 
@@ -78,6 +80,7 @@ jobs:
     if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') &&
       github.event.pull_request.draft == false
     strategy:
+      fail-fast: false
       matrix:
         runner: [linux-mi300-gpu-1, linux-mi325-gpu-1]
     runs-on: ${{matrix.runner}}
@@ -102,6 +105,7 @@ jobs:
     if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') &&
         github.event.pull_request.draft == false
     strategy:
+      fail-fast: false
       matrix:
         runner: [linux-mi300-gpu-1, linux-mi325-gpu-1]
     runs-on: ${{matrix.runner}}
@@ -142,6 +146,7 @@ jobs:
     if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') &&
         github.event.pull_request.draft == false
     strategy:
+      fail-fast: false
       matrix:
         runner: [linux-mi300-gpu-1, linux-mi325-gpu-1]
     runs-on: ${{matrix.runner}}
@@ -176,6 +181,7 @@ jobs:
     if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') &&
       github.event.pull_request.draft == false
     strategy:
+      fail-fast: false
       matrix:
         runner: [linux-mi300-gpu-2, linux-mi325-gpu-2]
     runs-on: ${{matrix.runner}}
@@ -242,12 +248,13 @@ jobs:
         run: |
           bash scripts/ci/amd_ci_exec.sh python3 run_suite.py --suite per-commit-amd --auto-partition-id ${{ matrix.part }} --auto-partition-size 8
 
-  unit-test-backend-2-gpu-amd:
+  unit-test-backend-1-gpu-amd-mi35x:
     if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') &&
       github.event.pull_request.draft == false
     strategy:
+      fail-fast: false
       matrix:
-        runner: [linux-mi300-gpu-2, linux-mi325-gpu-2]
+        runner: [linux-mi35x-gpu-1]
     runs-on: ${{matrix.runner}}
     steps:
       - name: Checkout code
@@ -262,16 +269,17 @@ jobs:
         run: bash scripts/ci/amd_ci_install_dependency.sh
 
       - name: Run test
-        timeout-minutes: 40
+        timeout-minutes: 50
         run: |
-          bash scripts/ci/amd_ci_exec.sh python3 run_suite.py --suite per-commit-2-gpu-amd
+          bash scripts/ci/amd_ci_exec.sh python3 run_suite.py --suite per-commit-amd-mi35x
 
-  unit-test-backend-8-gpu-amd:
+  unit-test-backend-2-gpu-amd:
     if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') &&
       github.event.pull_request.draft == false
     strategy:
+      fail-fast: false
       matrix:
-        runner: [linux-mi300-gpu-8]
+        runner: [linux-mi300-gpu-2, linux-mi325-gpu-2]
     runs-on: ${{matrix.runner}}
     steps:
       - name: Checkout code
@@ -286,14 +294,15 @@ jobs:
         run: bash scripts/ci/amd_ci_install_dependency.sh
 
       - name: Run test
-        timeout-minutes: 60
+        timeout-minutes: 40
         run: |
-          bash scripts/ci/amd_ci_exec.sh python3 run_suite.py --suite per-commit-8-gpu-amd --timeout-per-file 3600
+          bash scripts/ci/amd_ci_exec.sh python3 run_suite.py --suite per-commit-2-gpu-amd
 
-  unit-test-backend-8-gpu-CAR-amd:
+  unit-test-backend-8-gpu-amd:
     if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') &&
       github.event.pull_request.draft == false
     strategy:
+      fail-fast: false
       matrix:
         runner: [linux-mi300-gpu-8]
     runs-on: ${{matrix.runner}}
@@ -309,10 +318,10 @@ jobs:
       - name: Install dependencies
         run: bash scripts/ci/amd_ci_install_dependency.sh
 
-      - name: Run CustomAllReduce test
-        timeout-minutes: 20
+      - name: Run test
+        timeout-minutes: 60
         run: |
-          bash scripts/ci/amd_ci_exec.sh -e CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python3 -m unittest test_custom_allreduce.TestCustomAllReduce
+          bash scripts/ci/amd_ci_exec.sh python3 run_suite.py --suite per-commit-8-gpu-amd --timeout-per-file 3600
 
   unit-test-sgl-kernel-amd:
     if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') &&
@@ -350,8 +359,8 @@ jobs:
     needs: [
       accuracy-test-1-gpu-amd, mla-test-1-gpu-amd, bench-test-2-gpu-amd,
       accuracy-test-2-gpu-amd, performance-test-1-gpu-part-1-amd, performance-test-1-gpu-part-2-amd,
-      unit-test-backend-1-gpu-amd, unit-test-backend-2-gpu-amd, unit-test-backend-8-gpu-amd,
-      unit-test-sgl-kernel-amd
+      unit-test-backend-1-gpu-amd, unit-test-backend-1-gpu-amd-mi35x, unit-test-backend-2-gpu-amd,
+      unit-test-backend-8-gpu-amd, unit-test-sgl-kernel-amd
     ]
     runs-on: ubuntu-latest
     steps:

python/sglang/srt/models/deepseek_v2.py

@@ -2027,7 +2027,10 @@ class DeepseekV2DecoderLayer(nn.Module):
         quant_format = (
             "mxfp4"
             if _is_gfx95_supported
-            and self.self_attn.fused_qkv_a_proj_with_mqa.weight == torch.uint8
+            and getattr(self.self_attn, "fused_qkv_a_proj_with_mqa", None) is not None
+            and getattr(self.self_attn.fused_qkv_a_proj_with_mqa, "weight", None)
+            is not None
+            and self.self_attn.fused_qkv_a_proj_with_mqa.weight.dtype == torch.uint8
             else ""
         )
 
@@ -2582,7 +2585,11 @@ class DeepseekV2ForCausalLM(nn.Module):
                 0, (-1, self_attn.qk_nope_head_dim + self_attn.v_head_dim)
             ).split([self_attn.qk_nope_head_dim, self_attn.v_head_dim], dim=1)
 
-            if _use_aiter_gfx95 and self.quant_config.get_name() == "quark":
+            if (
+                _use_aiter_gfx95
+                and self.quant_config is not None
+                and self.quant_config.get_name() == "quark"
+            ):
                 w_kc, self_attn.w_scale_k, w_vc, self_attn.w_scale_v = (
                     quark_post_load_weights(self_attn, w, "mxfp4")
                 )

scripts/ci/amd_ci_exec.sh

 #!/bin/bash
 set -euo pipefail
 
+# Detect GPU family from hostname (e.g., linux-mi35x-gpu-1-xxxxx-runner-zzzzz)
+HOSTNAME_VALUE=$(hostname)
+GPU_FAMILY=""
+
+# Host names look like: linux-mi35x-gpu-1-xxxxx-runner-zzzzz
+if [[ "${HOSTNAME_VALUE}" =~ ^linux-(mi[0-9]+[a-z]*)-gpu-[0-9]+ ]]; then
+  GPU_FAMILY="${BASH_REMATCH[1]}"
+  echo "Detected GPU family from hostname: ${GPU_FAMILY}"
+else
+  echo "Warning: could not parse GPU family from '${HOSTNAME_VALUE}'"
+fi
+
 WORKDIR="/sglang-checkout/test/srt"
 declare -A ENV_MAP=(
   [SGLANG_AMD_CI]=1
@@ -8,6 +20,11 @@ declare -A ENV_MAP=(
   [SGLANG_USE_AITER]=1
 )
 
+# Conditionally add GPU_ARCHS only for mi35x
+if [[ "${GPU_FAMILY}" == "mi35x" ]]; then
+  ENV_MAP[GPU_ARCHS]="gfx950"
+fi
+
 # Parse -w/--workdir and -e ENV=VAL
 while [[ $# -gt 0 ]]; do
   case "$1" in

scripts/ci/amd_ci_install_dependency.sh

 #!/bin/bash
 set -euo pipefail
+HOSTNAME_VALUE=$(hostname)
+GPU_ARCH="mi30x"   # default
+
+# Host names look like: linux-mi35x-gpu-1-xxxxx-runner-zzzzz
+if [[ "${HOSTNAME_VALUE}" =~ ^linux-(mi[0-9]+[a-z]*)-gpu-[0-9]+ ]]; then
+  GPU_ARCH="${BASH_REMATCH[1]}"
+  echo "Detected GPU architecture from hostname: ${GPU_ARCH}"
+else
+  echo "Warning: could not parse GPU architecture from '${HOSTNAME_VALUE}', defaulting to ${GPU_ARCH}"
+fi
 
 # Install the required dependencies in CI.
 docker exec ci_sglang pip install --upgrade pip
 docker exec ci_sglang pip uninstall sgl-kernel -y || true
 docker exec -w /sglang-checkout/sgl-kernel ci_sglang bash -c "rm -f pyproject.toml && mv pyproject_rocm.toml pyproject.toml && python3 setup_rocm.py install"
-docker exec ci_sglang pip install -e "python[dev_hip]"
+
+case "${GPU_ARCH}" in
+  mi35x)
+    echo "Runner uses ${GPU_ARCH}; will fetch mi35x image."
+    docker exec ci_sglang pip install -e "python[dev_hip]" --no-deps # TODO: only for mi35x
+    # For lmms_evals evaluating MMMU
+    docker exec -w / ci_sglang git clone --branch v0.3.3 --depth 1 https://github.com/EvolvingLMMs-Lab/lmms-eval.git
+    docker exec -w /lmms-eval ci_sglang pip install -e . --no-deps # TODO: only for mi35x
+    ;;
+  mi30x|mi300|mi325)
+    echo "Runner uses ${GPU_ARCH}; will fetch mi30x image."
+    docker exec ci_sglang pip install -e "python[dev_hip]"
+    # For lmms_evals evaluating MMMU
+    docker exec -w / ci_sglang git clone --branch v0.3.3 --depth 1 https://github.com/EvolvingLMMs-Lab/lmms-eval.git
+    docker exec -w /lmms-eval ci_sglang pip install -e .
+    ;;
+  *)
+    echo "Runner architecture '${GPU_ARCH}' unrecognised;" >&2
+    ;;
+esac
 
 docker exec -w / ci_sglang git clone https://github.com/merrymercy/human-eval.git
 docker exec -w /human-eval ci_sglang pip install -e .
 
-# For lmms_evals evaluating MMMU
-docker exec -w / ci_sglang git clone --branch v0.3.3 --depth 1 https://github.com/EvolvingLMMs-Lab/lmms-eval.git
-docker exec -w /lmms-eval ci_sglang pip install -e .
-
 docker exec -w / ci_sglang mkdir -p /dummy-grok
 mkdir -p dummy-grok && wget https://sharkpublic.blob.core.windows.net/sharkpublic/sglang/dummy_grok.json -O dummy-grok/config.json
 docker cp ./dummy-grok ci_sglang:/

scripts/ci/amd_ci_start_container.sh

@@ -25,130 +25,102 @@ else
   echo "Warning: version.py not found, using default version: $SGLANG_VERSION" >&2
 fi
 
+
 # Default base tags (can be overridden by command line arguments)
 DEFAULT_MI30X_BASE_TAG="${SGLANG_VERSION}-rocm630-mi30x"
 DEFAULT_MI35X_BASE_TAG="${SGLANG_VERSION}-rocm700-mi35x"
 
 # Parse command line arguments
-MI30X_BASE_TAG="$DEFAULT_MI30X_BASE_TAG"
-MI35X_BASE_TAG="$DEFAULT_MI35X_BASE_TAG"
+MI30X_BASE_TAG="${DEFAULT_MI30X_BASE_TAG}"
+MI35X_BASE_TAG="${DEFAULT_MI35X_BASE_TAG}"
 
 while [[ $# -gt 0 ]]; do
   case $1 in
-    --mi30x-base-tag)
-      MI30X_BASE_TAG="$2"
-      shift 2
-      ;;
-    --mi35x-base-tag)
-      MI35X_BASE_TAG="$2"
-      shift 2
-      ;;
+    --mi30x-base-tag) MI30X_BASE_TAG="$2"; shift 2;;
+    --mi35x-base-tag) MI35X_BASE_TAG="$2"; shift 2;;
     -h|--help)
       echo "Usage: $0 [--mi30x-base-tag TAG] [--mi35x-base-tag TAG]"
-      echo "  --mi30x-base-tag TAG    Base tag for mi30x images (default: $DEFAULT_MI30X_BASE_TAG)"
-      echo "  --mi35x-base-tag TAG    Base tag for mi35x images (default: $DEFAULT_MI35X_BASE_TAG)"
       exit 0
       ;;
-    *)
-      echo "Unknown option $1"
-      echo "Use --help for usage information"
-      exit 1
-      ;;
+    *) echo "Unknown option $1"; exit 1;;
   esac
 done
 
+
+
+# Detect GPU architecture from the Kubernetes runner hostname
+HOSTNAME_VALUE=$(hostname)
+GPU_ARCH="mi30x"   # default
+
+# Host names look like: linux-mi35x-gpu-1-xxxxx-runner-zzzzz
+if [[ "${HOSTNAME_VALUE}" =~ ^linux-(mi[0-9]+[a-z]*)-gpu-[0-9]+ ]]; then
+  GPU_ARCH="${BASH_REMATCH[1]}"
+  echo "Detected GPU architecture from hostname: ${GPU_ARCH}"
+else
+  echo "Warning: could not parse GPU architecture from '${HOSTNAME_VALUE}', defaulting to ${GPU_ARCH}"
+fi
+
+# Normalise / collapse architectures we don’t yet build specifically for
+case "${GPU_ARCH}" in
+  mi35x)
+    echo "Runner uses ${GPU_ARCH}; will fetch mi35x image."
+    ;;
+  mi30x|mi300|mi325)
+    echo "Runner uses ${GPU_ARCH}; will fetch mi30x image."
+    GPU_ARCH="mi30x"
+    ;;
+  *)
+    echo "Runner architecture '${GPU_ARCH}' unrecognised; defaulting to mi30x image." >&2
+    GPU_ARCH="mi30x"
+    ;;
+esac
+
+
 # Set up DEVICE_FLAG based on Kubernetes pod info
-if [ -f "/etc/podinfo/gha-render-devices" ]; then
+if [[ -f /etc/podinfo/gha-render-devices ]]; then
   DEVICE_FLAG=$(cat /etc/podinfo/gha-render-devices)
 else
   DEVICE_FLAG="--device /dev/dri"
 fi
 
 
-
-# Function to find latest available image for a given GPU architecture
+# Find the latest image
 find_latest_image() {
   local gpu_arch=$1
-  local base_tag
-
-  if [ "$gpu_arch" == "mi30x" ]; then
-    base_tag="$MI30X_BASE_TAG"
-  elif [ "$gpu_arch" == "mi35x" ]; then
-    base_tag="$MI35X_BASE_TAG"
-  else
-    echo "Error: Unsupported GPU architecture '$gpu_arch'" >&2
-    return 1
-  fi
-
-  local days_back=0
+  local base_tag days_back image_tag
 
-  while [ $days_back -lt 7 ]; do
-    local check_date=$(date -d "$days_back days ago" +%Y%m%d)
-    local image_tag="${base_tag}-${check_date}"
+  case "${gpu_arch}" in
+      mi30x) base_tag="${MI30X_BASE_TAG}" ;;
+      mi35x) base_tag="${MI35X_BASE_TAG}" ;;
+      *)     echo "Error: unsupported GPU architecture '${gpu_arch}'" >&2; return 1 ;;
+  esac
 
+  for days_back in {0..6}; do
+    image_tag="${base_tag}-$(date -d "${days_back} days ago" +%Y%m%d)"
     echo "Checking for image: rocm/sgl-dev:${image_tag}" >&2
-
-    # Check if the image exists by trying to get its manifest
     if docker manifest inspect "rocm/sgl-dev:${image_tag}" >/dev/null 2>&1; then
       echo "Found available image: rocm/sgl-dev:${image_tag}" >&2
       echo "rocm/sgl-dev:${image_tag}"
       return 0
     fi
-
-    days_back=$((days_back + 1))
   done
 
-  echo "Error: No ${gpu_arch} image found in the last 7 days for version ${base_tag}" >&2
-
-  # Final fallback to specific hardcoded images
-  echo "Using final fallback images..." >&2
-  if [ "$gpu_arch" == "mi30x" ]; then
-    echo "rocm/sgl-dev:v0.5.0rc0-rocm630-mi30x-20250812"
-  elif [ "$gpu_arch" == "mi35x" ]; then
+  echo "Error: no ${gpu_arch} image found in the last 7 days for base ${base_tag}" >&2
+  echo "Using hard-coded fallback…" >&2
+  if [[ "${gpu_arch}" == "mi35x" ]]; then
     echo "rocm/sgl-dev:v0.5.0rc0-rocm700-mi35x-20250812"
   else
-    echo "rocm/sgl-dev:v0.5.0rc0-rocm630-mi30x-20250812"  # Default to mi30x
+    echo "rocm/sgl-dev:v0.5.0rc0-rocm630-mi30x-20250812"
   fi
-
-  return 0
 }
 
-# Determine image finder and fallback based on runner
-# In Kubernetes, the hostname contains the GPU type (e.g., linux-mi300-gpu-1-bgg8r-runner-vknlb)
-# Extract the GPU type from hostname
-HOSTNAME_VALUE=$(hostname)
-RUNNER_NAME="unknown"
-
-if [[ "${HOSTNAME_VALUE}" =~ ^(linux-mi[0-9]+-gpu-[0-9]+) ]]; then
-  RUNNER_NAME="${BASH_REMATCH[1]}"
-  echo "Extracted runner from hostname: ${RUNNER_NAME}"
-else
-  echo "Could not extract runner info from hostname: ${HOSTNAME_VALUE}"
-fi
-
-echo "The runner is: ${RUNNER_NAME}"
-GPU_ARCH="mi30x"
-
-# Check for mi350/mi355 runners
-if [[ "${RUNNER_NAME}" =~ ^linux-mi350-gpu-[0-9]+$ ]] || [[ "${RUNNER_NAME}" =~ ^linux-mi355-gpu-[0-9]+$ ]]; then
-  echo "Runner is ${RUNNER_NAME}, will find mi35x image."
-  GPU_ARCH="mi35x"
-# Check for mi300/mi325 runners
-elif [[ "${RUNNER_NAME}" =~ ^linux-mi300-gpu-[0-9]+$ ]] || [[ "${RUNNER_NAME}" =~ ^linux-mi325-gpu-[0-9]+$ ]]; then
-  echo "Runner is ${RUNNER_NAME}, will find mi30x image."
-else
-  echo "Runner type not recognized: '${RUNNER_NAME}'"
-  echo "Defaulting to find mi30x image"
-fi
-
-# Find and pull the latest image
+# Pull and run the latest image
 IMAGE=$(find_latest_image "${GPU_ARCH}")
-echo "Pulling Docker image: $IMAGE"
-docker pull "$IMAGE"
+echo "Pulling Docker image: ${IMAGE}"
+docker pull "${IMAGE}"
 
-# Run the container
-echo "Starting container: ci_sglang"
-docker run -dt --user root --device=/dev/kfd $DEVICE_FLAG \
+echo "Launching container: ci_sglang"
+docker run -dt --user root --device=/dev/kfd ${DEVICE_FLAG} \
   -v "${GITHUB_WORKSPACE:-$PWD}:/sglang-checkout" \
   --ipc=host --group-add video \
   --shm-size 32g \
@@ -157,4 +129,4 @@ docker run -dt --user root --device=/dev/kfd $DEVICE_FLAG \
   --security-opt seccomp=unconfined \
   -w /sglang-checkout \
   --name ci_sglang \
-  "$IMAGE"
+  "${IMAGE}"

test/srt/run_suite.py

@@ -243,6 +243,10 @@ suite_amd = {
         TestFile("test_wave_attention_kernels.py", 2),
         TestFile("test_wave_attention_backend.py", 150),
     ],
+    "per-commit-amd-mi35x": [
+        TestFile("test_mla.py", 242),
+        TestFile("test_gpt_oss_1gpu.py", 600),
+    ],
     "per-commit-2-gpu-amd": [
         TestFile("lora/test_lora_tp.py", 116),
         TestFile("rl/test_update_weights_from_distributed.py", 103),

test/srt/test_gpt_oss_common.py

+import os
 from concurrent.futures import ThreadPoolExecutor
 from types import SimpleNamespace
 from typing import Dict, List, Literal, Optional
 
-from sglang.srt.utils import kill_process_tree
+from sglang.srt.utils import is_hip, kill_process_tree
 from sglang.test.run_eval import run_eval
 from sglang.test.test_utils import (
     DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
@@ -14,6 +15,7 @@ from sglang.test.test_utils import (
 )
 
 _base_url = DEFAULT_URL_FOR_TEST
+_is_hip = is_hip()
 
 
 class BaseTestGptOss(CustomTestCase):
@@ -36,7 +38,8 @@ class BaseTestGptOss(CustomTestCase):
 
         if model_variant == "20b":
             other_args += ["--cuda-graph-max-bs", "600"]
-
+        if _is_hip:
+            os.environ["SGLANG_USE_AITER"] = "0"
         self._run_test_raw(
             model=model,
             expected_score_of_reasoning_effort=expected_score_of_reasoning_effort,