Migrate H100/A100 runners to OSDC (ARC) (#179)

* Migrate linux.aws.h100/a100 runners to OSDC (ARC) runners

Switch all linux.aws.h100* and linux.aws.a100 runner labels to their
OSDC/ARC equivalents. Labels follow the mapping in pytorch/pytorch
.github/arc.yaml, with the mt- (Meta multi-tenant) prefix that OSDC
production runners use:

  linux.aws.a100   -> mt-l-x86iavx512-11-125-a100   (1 GPU)
  linux.aws.h100   -> mt-l-x86iamx-22-225-h100       (1 GPU)
  linux.aws.h100.4 -> mt-l-x86iamx-88-900-h100-4     (4 GPU)
  linux.aws.h100.8 -> mt-l-bx86iamx-176-1800-h100-8  (8 GPU)

Files:
- generate_vllm_benchmark_matrix.py: TP_TO_RUNNER_MAPPING and
  RUNNER_TO_PLATFORM_MAPPING get the full label rename. In PLATFORM_SKIPS
  the skip tokens become the bare GPU-type 'h100'/'a100' so they remain a
  substring of the OSDC names, preserving the 'skip the whole family'
  behavior the substring matcher relies on (matters for h100, which has
  1/4/8-GPU variants).
- vllm-ci-test.yml, vllm-profiling.yml, pytorch-bisect.yaml: runs-on /
  runner choice updated.
- test fixture: expected runner values updated to the OSDC names.

The matrix output is unchanged except for the runner label strings
(verified: every model<->runner pairing is identical after the rename).

* Containerize vLLM CI test + profiling workflows for OSDC runners

OSDC/ARC runners are ephemeral pods with no Docker daemon, so the old
'docker run --gpus all + docker exec' pattern cannot work on them. Run the
vLLM CI image via the job-level container: key with options '--gpus all'
instead (the GPU is injected by the runner pod), matching pytorch/pytorch
_linux-test.yml (test-osdc) and pytorch/helion.

- vllm-ci-test.yml / vllm-profiling.yml: add an ubuntu-latest 'resolve-image'
  pre-job that runs 'docker manifest inspect' (needs a daemon the pod lacks)
  to pick the latest available vLLM CI image and pass it down as the
  container image. Drop the GPU_FLAG/docker run/docker exec wrapper and the
  /tmp/workspace bind-mount; run the scripts directly in the container.
- vllm-profiling.yml: assume the upload IAM role via OIDC before the S3
  upload (ephemeral pods have no host instance role); pass the resolved
  vLLM commit through as S3_HEAD_SHA.
- run_vllm_profiling.sh: use $GITHUB_WORKSPACE instead of the hardcoded
  /tmp/workspace bind-mount path.

* Containerize pytorch-bisect.yaml mt- (OSDC) path in a CUDA devel container

The mt- runner is an ephemeral OSDC pod with no host CUDA toolchain, so build
PyTorch inside pytorch/pytorch:2.12.0-cuda13.0-cudnn9-devel (--gpus all) when
the mt- runner is selected; linux.dgx.b200 keeps the existing bare-host path
(conditional container via fromJSON('null')). CUDA_HOME points at the image's
/usr/local/cuda on the container path (run.sh requires it non-empty). Add a
git safe.directory step for the root-owned in-container checkout.

* Containerize vllm-benchmark + sglang benchmarks for OSDC (all devices)

Run every matrix device inside a job-level container: instead of the old
'docker run + docker exec' pattern, since OSDC/ARC pods have no Docker daemon.

- generate_vllm_benchmark_matrix.py: emit a per-entry 'device-name' so the
  workflow can resolve the container image up front (regenerated the test
  fixture, which also clears pre-existing config drift).
- set-parameters: resolve the upstream image on ubuntu-latest (which has a
  daemon) via 'docker manifest inspect', then enrich every matrix entry with
  container-image + device-appropriate container-options. sglang resolves per
  image suffix (cuda / -cu128-b200 / -rocm630-mi30x) and skips non-cuda/rocm
  devices instead of failing the whole matrix.
- benchmarks job: add container: { image, options }, drop the device probe
  (device-name comes from the matrix) while keeping the runtime DEVICE_TYPE
  detection, run the benchmark script natively, and assume the upload IAM
  role via OIDC for all devices (no host instance role inside a pod).
  chown is made sudo-optional for the in-container root user.

Flagged for CI validation: the per-model S3 'already benchmarked' dedup is
dropped (needs the runtime device-type before the container exists); the
rocm/hpu container options are best-effort; and --shm-size may be capped on
ARC pods.

* Revert benchmark/sglang containerization; keep migration minimal

The H100/A100 benchmark jobs do not actually run in this repo - they run
in pytorch/pytorch's .github/workflows/vllm-benchmark.yml, which consumes
the runner labels emitted by generate_vllm_benchmark_matrix.py. So this
repo's vllm-benchmark.yml / sglang-benchmark.yml never execute on the
migrated OSDC runners and do not need the job-level container rewrite.

Revert f6bc62f's four-file change:
- vllm-benchmark.yml / sglang-benchmark.yml: back to the docker run/exec
  pattern (they run on b200 / rocm / cpu / hpu, all of which keep a Docker
  daemon - not OSDC).
- generate_vllm_benchmark_matrix.py: drop the per-entry device-name field
  (it only existed to drive the dropped containerization). The H100/A100
  -> OSDC label rename is kept, since pytorch/pytorch needs it.
- test fixture: regenerated from the current benchmark configs (clears
  pre-existing config drift) with the mt- labels and no device-name.

What remains is the minimal migration: the runner label rename plus the
containerization of the three workflows that DO run on the migrated
runners here (vllm-ci-test, vllm-profiling, pytorch-bisect).

Author: huydhn

.github/scripts/generate_vllm_benchmark_matrix.py

@@ -14,7 +14,7 @@
 # while ROCm runner are provided by AMD
 TP_TO_RUNNER_MAPPING = {
     1: [
-        "linux.aws.h100",
+        "mt-l-x86iamx-22-225-h100",
         "linux.rocm.gpu.gfx942.1",
         "linux.24xl.spr-metal",
         "linux.24xl.gnr",
@@ -25,20 +25,20 @@
     # NB: There is no 2xH100 runner at the momement, so let's use the next one
     # in the list here which is 4xH100
     2: [
-        "linux.aws.h100.4",
+        "mt-l-x86iamx-88-900-h100-4",
         "linux.rocm.gpu.gfx942.2",
         "linux.24xl.gnr",
         "linux.hpu.gaudi3.8",
         "linux.dgx.b200.8",
     ],
     4: [
-        "linux.aws.h100.4",
+        "mt-l-x86iamx-88-900-h100-4",
         "linux.rocm.gpu.gfx942.4",
         "linux.hpu.gaudi3.8",
         "linux.dgx.b200.8",
     ],
     8: [
-        "linux.aws.h100.8",
+        "mt-l-bx86iamx-176-1800-h100-8",
         "linux.rocm.gpu.gfx942.8",
         "linux.dgx.b200.8",
         "linux.hpu.gaudi3.8",
@@ -47,10 +47,10 @@
 
 # This mapping is needed to find out the platform of the runner
 RUNNER_TO_PLATFORM_MAPPING = {
-    "linux.aws.a100": "cuda",
-    "linux.aws.h100": "cuda",
-    "linux.aws.h100.4": "cuda",
-    "linux.aws.h100.8": "cuda",
+    "mt-l-x86iavx512-11-125-a100": "cuda",
+    "mt-l-x86iamx-22-225-h100": "cuda",
+    "mt-l-x86iamx-88-900-h100-4": "cuda",
+    "mt-l-bx86iamx-176-1800-h100-8": "cuda",
     "linux.dgx.b200": "cuda",
     "linux.dgx.b200.8": "cuda",
     "linux.rocm.gpu.gfx942.1": "rocm",
@@ -102,61 +102,61 @@
     ],
     # Run some bigger models on B200 to share the load
     "Qwen/Qwen3-30B-A3B": [
-        "linux.aws.a100",
-        "linux.aws.h100",
+        "a100",
+        "h100",
         "linux.rocm.gpu.gfx942",  # TODO: Fail on ROCm
         "linux.24xl.gnr",
         "linux.hpu.gaudi3.8",
     ],
     "google/gemma-3-27b-it": [
-        "linux.aws.a100",
-        "linux.aws.h100",
+        "a100",
+        "h100",
         "linux.rocm.gpu.gfx942",  # TODO (huydhn): Fail on ROCm
         "linux.24xl.gnr",
         "linux.hpu.gaudi3.8",
     ],
     "meta-llama/Llama-4-Scout-17B-16E-Instruct": [
-        "linux.aws.a100",
-        "linux.aws.h100",
+        "a100",
+        "h100",
         "linux.rocm.gpu.gfx942",  # TODO: Fail on ROCm
         "linux.24xl.gnr",
         "linux.hpu.gaudi3.8",
     ],
     "meta-llama/Llama-4-Maverick-17B-128E-Instruct-FP8": [
-        "linux.aws.a100",
-        "linux.aws.h100",
+        "a100",
+        "h100",
         "linux.rocm.gpu.gfx942",  # TODO (huydhn): Hang on ROCm
         "linux.24xl.gnr",
         "linux.hpu.gaudi3.8",
     ],
     # Run gpt-oss on both H100 and B200
     "openai/gpt-oss-20b": [
-        "linux.aws.a100",
+        "a100",
         "linux.24xl.gnr",
         "linux.hpu.gaudi3.8",
     ],
     "openai/gpt-oss-120b": [
-        "linux.aws.a100",
+        "a100",
         "linux.24xl.gnr",
         "linux.hpu.gaudi3.8",
     ],
     # Deepseek can only run on B200
     "deepseek-ai/DeepSeek-V3.1": [
-        "linux.aws.a100",
-        "linux.aws.h100",
+        "a100",
+        "h100",
         "linux.24xl.gnr",
         "linux.hpu.gaudi3.8",
     ],
     "deepseek-ai/DeepSeek-V3.2": [
-        "linux.aws.a100",
-        "linux.aws.h100",
+        "a100",
+        "h100",
         "linux.24xl.gnr",
         "linux.hpu.gaudi3.8",
     ],
     "deepseek-ai/DeepSeek-R1": [
-        "linux.aws.a100",
+        "a100",
         "linux.24xl.gnr",
-        "linux.aws.h100",
+        "h100",
         "linux.hpu.gaudi3.8",
     ],
 }

.github/scripts/run_vllm_profiling.sh

@@ -12,7 +12,9 @@ print_configuration() {
 }
 
 setup_workspace() {
-    WORKSPACE_DIR="/tmp/workspace"
+    # On OSDC/ARC the job runs inside a container and the repo is checked out at
+    # $GITHUB_WORKSPACE rather than the old docker-run bind mount at /tmp/workspace.
+    WORKSPACE_DIR="${GITHUB_WORKSPACE:-/tmp/workspace}"
     cd "${WORKSPACE_DIR}"
 
     echo "Creating profiling directory: ${VLLM_TORCH_PROFILER_DIR}"