[AMD] docker: retire the rocm700 variants, add a rocm10-mi35x variant

.github/workflows/docker-build.yml

@@ -25,9 +25,8 @@ on:
           - cu13-x86
           - cu13-aarch64
           - cu12-x86
-          - rocm700-mi35x
-          - rocm700-mi30x
           - rocm720-mi35x
+          - rocm10-mi35x
       image_tag:
         description: 'Tag mode: latest, dev (rolling + timestamped), or custom'
         required: true

docker/Dockerfile.rocm

@@ -1,17 +1,16 @@
 # doc-dev: docs/ci/02-docker-build.md
-# Multi-arch ROCm image for Miles (built on a prebuilt sglang base):
+# ROCm / gfx950 image for Miles (built on a prebuilt sglang base):
 #
-# 1. rocm700-mi35x  (gfx950, ROCm 7.0 / MI350 / MI355X)
-#    build-arg:SGLANG_IMAGE_REPO=rocm/sgl-dev  SGLANG_IMAGE_TAG=v0.5.14-rocm700-mi35x-20260627
-#
-# 2. rocm700-mi30x  (gfx942, ROCm 7.0 / MI300 / MI325)
-#    build-arg:GPU_ARCH=gfx942  SGLANG_IMAGE_TAG=v0.5.10-rocm700-mi30x
-#
-# 3. rocm720-mi35x  (gfx950, ROCm 7.2 / MI350 / MI355X)
+# 1. rocm720-mi35x  (gfx950, ROCm 7.2 / MI350 / MI355X, Python 3.10 base)
 #    build-arg:SGLANG_IMAGE_REPO=rocm/sgl-dev  SGLANG_IMAGE_TAG=v0.5.16-rocm720-mi35x-20260730
+#                WHEELS_TAG_ROCM=rocm720-gfx950-v0.5.16  TE_USE_WHEEL=1  APPLY_ROCR_VMMFIX=1
+#
+# 2. rocm10-mi35x   (gfx950, ROCm 10 / MI350 / MI355X, Python 3.12 base)
+#    build-arg:SGLANG_IMAGE_REPO=rocm/sgl-dev  SGLANG_IMAGE_TAG=v0.5.18-rocm10-mi35x-20260831
+#                WHEELS_TAG_ROCM=rocm10-gfx950-v0.5.18  TE_USE_WHEEL=1   (cp312 shelf)
 
-ARG SGLANG_IMAGE_REPO=lmsysorg/sglang
-ARG SGLANG_IMAGE_TAG=v0.5.10-rocm700-mi35x
+ARG SGLANG_IMAGE_REPO=rocm/sgl-dev
+ARG SGLANG_IMAGE_TAG=v0.5.16-rocm720-mi35x-20260730
 FROM ${SGLANG_IMAGE_REPO}:${SGLANG_IMAGE_TAG} AS sglang
 
 SHELL ["/bin/bash", "-ceuxo", "pipefail"]
@@ -26,9 +25,8 @@ ARG MEGATRON_BRANCH=miles-main
 ARG MILES_COMMIT=main
 
 ARG GPU_ARCH=gfx950
-# 1 on ROCm 7.2 to bake the ROCr VMM-pause fix; ROCm 7.0 has no such regression, leave 0.
+# 1 bakes the rebuilt ROCr with the VMM-pause fix (ROCm 7.2 needs it; ROCm 10 has it upstream).
 ARG APPLY_ROCR_VMMFIX=0
-ARG SGLANG_USE_ROCM700A=0
 ARG WHEELS_REPO=XinyuJiangCMU/miles-wheels-rocm
 ARG WHEELS_TAG_ROCM=rocm720-gfx950-v0.5.16
 ARG MAX_JOBS=128
@@ -48,13 +46,13 @@ WORKDIR /root/
 
 ENV MAX_JOBS=${MAX_JOBS}
 
-# Build configuration for the selected arch (gfx942 / gfx950).
+# Build configuration for the target arch.
 ENV GPU_ARCH=${GPU_ARCH}
 ENV PYTORCH_ROCM_ARCH=${GPU_ARCH}
 ENV GPU_ARCH_LIST=${GPU_ARCH}
 ENV AMDGPU_TARGET=${GPU_ARCH}
 
-# Transformer Engine build knobs for the v2.8_rocm branch.
+# Transformer Engine build knobs (consumed by the TE_USE_WHEEL=0 source build).
 ENV NVTE_FRAMEWORK=pytorch
 ENV NVTE_ROCM_ARCH=${GPU_ARCH}
 ENV NVTE_USE_HIPBLASLT=1
@@ -89,7 +87,7 @@ RUN mkdir -p /tmp/wheels && \
     curl -fsSL "https://api.github.com/repos/${WHEELS_REPO}/releases/tags/${WHEELS_TAG_ROCM}" \
     | python3 -c "import sys,json,subprocess,os; w='/tmp/wheels'; [subprocess.run(['curl','-fsSL','-o',os.path.join(w,a['name']),a['browser_download_url']],check=True) for a in json.load(sys.stdin).get('assets',[]) if a['name'].endswith(('.whl','.tar.gz'))]" && \
     ls -lh /tmp/wheels/
-# Transformer Engine: prebuilt fp8 wheel
+# Transformer Engine: prebuilt wheel from the variant's wheels shelf, or source build
 ARG TE_USE_WHEEL=0
 RUN pip uninstall -y transformer-engine transformer_engine transformer_engine_torch || true; \
     if [ "${TE_USE_WHEEL}" = "1" ]; then \
@@ -113,7 +111,7 @@ RUN pip install flash-linear-attention==0.5.2
 # and fix TileLang's HIP fp8->float conversion to use ROCm device intrinsics instead of the SDK host-only operator.
 # Drop these patches once the corresponding tilelang/tile_kernels fixes land in the base image.
 RUN pip install --no-deps tile_kernels==1.0.0 && \
-    patch -p1 -d /opt/venv/lib/python3.10/site-packages < /tmp/amd_patch/tile_kernels.patch && \
+    patch -p1 -d "$(python3 -c 'import site; print(site.getsitepackages()[0])')" < /tmp/amd_patch/tile_kernels.patch && \
     patch -p1 -d /opt/tilelang/src/tl_templates/hip < /tmp/amd_patch/tilelang_hip_fp8.patch
 
 RUN rm -rf /root/Megatron-LM && \
@@ -136,7 +134,7 @@ RUN pip install "numba==0.65.1"
 
 RUN python -c "import sglang; import sgl_kernel; print('SGLang + sgl_kernel: OK')"
 
-RUN pip install git+https://github.com/fzyzcjy/torch_memory_saver.git@f05a8754daf68238d54e4cf31cb3ba866684bbaf --no-cache-dir --force-reinstall
+RUN pip install git+https://github.com/fzyzcjy/torch_memory_saver.git@f05a8754daf68238d54e4cf31cb3ba866684bbaf --no-cache-dir --force-reinstall --no-build-isolation
 
 # Fix the ROCm 7.2 ROCr VMM-pause leak.
 RUN if [ "$APPLY_ROCR_VMMFIX" = "1" ]; then \
@@ -151,7 +149,7 @@ RUN if [ "$APPLY_ROCR_VMMFIX" = "1" ]; then \
       install -m0644 "/tmp/$asset" "$dst" && rm -f "/tmp/$asset" && \
       echo "[vmmfix] installed patched ROCr at $dst"; \
     else \
-      echo "[vmmfix] skipped (APPLY_ROCR_VMMFIX=$APPLY_ROCR_VMMFIX; ROCm 7.0 has no VMM-pause regression)"; \
+      echo "[vmmfix] skipped (APPLY_ROCR_VMMFIX=$APPLY_ROCR_VMMFIX; ROCm 10 has the fix upstream)"; \
     fi
 RUN pip install git+https://github.com/radixark/Megatron-Bridge.git@bridge --no-deps --no-build-isolation
 RUN pip install megatron-energon --no-deps
@@ -170,9 +168,6 @@ RUN rm -rf /usr/lib/python3/dist-packages/jwt /usr/lib/python3/dist-packages/PyJ
 # Pin the CLI stack (typer/click) for huggingface_hub / Ray compatibility.
 RUN pip install "typer==0.25.1" "click==8.2.1"
 
-# Pin numpy 1.x for Megatron compatibility.
-RUN pip install "numpy<2"
-
 COPY docker/install-kube-tools.sh /tmp/install-kube-tools.sh
 RUN bash /tmp/install-kube-tools.sh "${TARGETARCH}" && rm -f /tmp/install-kube-tools.sh
 
@@ -226,7 +221,6 @@ ENV SGLANG_ALLOW_OVERWRITE_LONGER_CONTEXT_LEN=1
 ENV SGLANG_MOE_PADDING=1
 ENV SGLANG_SET_CPU_AFFINITY=1
 ENV SGLANG_ROCM_FUSED_DECODE_MLA=1
-ENV SGLANG_USE_ROCM700A=${SGLANG_USE_ROCM700A}
 ENV NCCL_MIN_NCHANNELS=112
 ENV TORCHINDUCTOR_MAX_AUTOTUNE=1
 ENV TORCHINDUCTOR_MAX_AUTOTUNE_POINTWISE=1

docker/build.py

@@ -50,39 +50,30 @@
             "WHEELS_TAG_X86": "cu129-x86_64",
         },
     },
-    "rocm700-mi35x": {
+    "rocm720-mi35x": {
         "image": "rocm/sgl-dev",
-        "tag_postfix": "-rocm700-mi35x",
+        "tag_postfix": "-rocm720-mi35x",
         "tag_prefix": "miles",
         "dockerfile": "docker/Dockerfile.rocm",
         "build_args": {
             "GPU_ARCH": "gfx950",
             "SGLANG_IMAGE_REPO": "rocm/sgl-dev",
-            "SGLANG_IMAGE_TAG": "v0.5.14-rocm700-mi35x-20260627",
-            "SGLANG_USE_ROCM700A": "1",
-        },
-    },
-    "rocm700-mi30x": {
-        "image": "rocm/sgl-dev",
-        "tag_postfix": "-rocm700-mi30x",
-        "tag_prefix": "miles",
-        "dockerfile": "docker/Dockerfile.rocm",
-        "build_args": {
-            "GPU_ARCH": "gfx942",
-            "SGLANG_IMAGE_TAG": "v0.5.10-rocm700-mi30x",
-            "SGLANG_USE_ROCM700A": "1",
+            "SGLANG_IMAGE_TAG": "v0.5.16-rocm720-mi35x-20260730",
+            "WHEELS_TAG_ROCM": "rocm720-gfx950-v0.5.16",
+            "APPLY_ROCR_VMMFIX": "1",
+            "TE_USE_WHEEL": "1",
         },
     },
-    "rocm720-mi35x": {
+    "rocm10-mi35x": {
         "image": "rocm/sgl-dev",
-        "tag_postfix": "-rocm720-mi35x",
+        "tag_postfix": "-rocm10-mi35x",
         "tag_prefix": "miles",
         "dockerfile": "docker/Dockerfile.rocm",
         "build_args": {
             "GPU_ARCH": "gfx950",
             "SGLANG_IMAGE_REPO": "rocm/sgl-dev",
-            "SGLANG_IMAGE_TAG": "v0.5.16-rocm720-mi35x-20260730",
-            "APPLY_ROCR_VMMFIX": "1",
+            "SGLANG_IMAGE_TAG": "v0.5.18-rocm10-mi35x-20260831",
+            "WHEELS_TAG_ROCM": "rocm10-gfx950-v0.5.18",
             "TE_USE_WHEEL": "1",
         },
     },
@@ -176,9 +167,8 @@ class Variant(str, Enum):
     cu13_x86 = "cu13-x86"
     cu13_aarch64 = "cu13-aarch64"
     cu12_x86 = "cu12-x86"
-    rocm700_mi35x = "rocm700-mi35x"
-    rocm700_mi30x = "rocm700-mi30x"
     rocm720_mi35x = "rocm720-mi35x"
+    rocm10_mi35x = "rocm10-mi35x"
 
 
 class ImageTag(str, Enum):

docs/ci/02-docker-build.md

@@ -10,7 +10,7 @@ GPU CI runs inside `radixark/miles`. This doc maps which Dockerfiles exist, the
 | Path                     | Builds                   | Wired into                             |
 | ------------------------ | ------------------------ | -------------------------------------- |
 | `docker/Dockerfile`      | `radixark/miles` (CUDA)  | `docker-build.yml`, `release-docker.yml` |
-| `docker/Dockerfile.rocm` | AMD ROCm (MI30x / MI35x) | `docker-build.yml` (`rocm7xx-mi3xx` variants) |
+| `docker/Dockerfile.rocm` | AMD ROCm (MI35x) | `docker-build.yml` (`rocm*-mi35x` variants) |
 
 
 ### `docker/Dockerfile` — inputs & output
@@ -31,7 +31,7 @@ The Dockerfile is the build recipe: it provides the cu13 defaults and emits one
 
 **Output** — one `radixark/miles` image for the platform buildx targets: the SGLang base, then the Python dependencies declared in `requirements.txt`, Megatron-LM at its branch default or caller-pinned commit, Miles, and the prebuilt wheels (`sgl-router` among them). A multi-arch build is one `buildx` run executed once per platform — `TARGETARCH` differs each time, so each arch installs its own wheels — and buildx pushes the two as a single manifest.
 
-`docker/Dockerfile.rocm` is the ROCm counterpart (build-args `GPU_ARCH` + a ROCm `SGLANG_IMAGE_TAG`; the 7.2 variants also set `APPLY_ROCR_VMMFIX=1`, which downloads the ROCr VMM-pause fix `.so` from the `WHEELS_TAG_ROCM` release and installs it — ROCm 7.0 has no such regression and leaves it off).
+`docker/Dockerfile.rocm` is the ROCm counterpart (build-args `GPU_ARCH` + a ROCm `SGLANG_IMAGE_TAG`). `rocm720-mi35x` sets `APPLY_ROCR_VMMFIX=1` to install the rebuilt ROCr with the VMM-pause fix from its wheels release; ROCm 10 has the fix upstream. `rocm10-mi35x` uses its own cp312 wheels release (`rocm10-gfx950-v0.5.18`).
 
 ## Build script
 
@@ -44,9 +44,8 @@ The Dockerfile is the build recipe: it provides the cu13 defaults and emits one
 | `cu13-x86`     | `radixark/miles:dev`               | `linux/amd64`                 | x86-only build of the same image               |
 | `cu13-aarch64` | `radixark/miles:dev`               | `linux/arm64`                 | arm64-only build of the same image             |
 | `cu12-x86`     | `radixark/miles:dev-cu12`          | `linux/amd64`                 | CUDA 12.9 legacy                               |
-| `rocm700-mi30x` | `rocm/sgl-dev:miles-rocm700-mi30x` | native                       | AMD MI30x (`gfx942`, ROCm 7.0) — `docker/Dockerfile.rocm` |
-| `rocm700-mi35x` | `rocm/sgl-dev:miles-rocm700-mi35x` | native                       | AMD MI35x (`gfx950`, ROCm 7.0) — `docker/Dockerfile.rocm` |
 | `rocm720-mi35x` | `rocm/sgl-dev:miles-rocm720-mi35x` | native                       | AMD MI35x (`gfx950`, ROCm 7.2) — `docker/Dockerfile.rocm` |
+| `rocm10-mi35x`  | `rocm/sgl-dev:miles-rocm10-mi35x`  | native                       | AMD MI35x (`gfx950`, ROCm 10, py3.12 base, ROCm SDK as pip wheels) — `docker/Dockerfile.rocm` |
 
 
 The cu13 variants share one multi-arch CUDA base image and differ only in platforms. `cu13` runs a single `buildx --platform linux/amd64,linux/arm64` — buildx builds both arches and pushes them as one manifest in a single shot, with the Dockerfile picking each layer's wheels by `TARGETARCH` (see Dockerfile inputs), so `docker pull` auto-selects by host arch.
@@ -86,7 +85,7 @@ This workflow owns the rolling `dev` and `latest` image families. It has two job
 ### Triggers: automatic vs manual
 
 - **Automatic** (no human) — the **schedule** (cron 00:00 / 12:00 UTC, gated by `check-upstream`) and any **push to `main` that touches `docker/Dockerfile`, `docker/install-kube-tools.sh`, `docker/verify_transformer_engine.py`, or `requirements.txt`**. Both leave `--variant` empty and build **two images**: `cu13` → `radixark/miles` (multi-arch) and `cu12-x86` → `radixark/miles:dev-cu12`.
-- **Manual** — `workflow_dispatch` (pick one variant — see Trigger a build yourself below) or running `docker/build.py` locally. Only the `rocm7xx-mi3xx` images have **no automatic path** (`cu13-x86` / `cu13-aarch64` just rebuild the same `dev` image single-arch).
+- **Manual** — `workflow_dispatch` (pick one variant — see Trigger a build yourself below) or running `docker/build.py` locally. Only the `rocm*-mi35x` images have **no automatic path in this repo** (their nightlies live in sgl-project/sglang) (`cu13-x86` / `cu13-aarch64` just rebuild the same `dev` image single-arch).
 
 `docker/build.py` and `docker/patch/**` participate in PR image validation but are not `main`-push triggers; [Release a Version](/ci/04-release) treats them as manual preflight cases.
 
@@ -107,7 +106,7 @@ All images push to **Docker Hub**. CUDA variants → `radixark/miles`; ROCm vari
 | --- | --- | --- |
 | `cu13` / `cu13-x86` / `cu13-aarch64` | `radixark/miles:dev` + `radixark/miles:dev-<YYYYMMDDHHMM>` | `radixark/miles:latest` |
 | `cu12-x86` | `radixark/miles:dev-cu12` (+ timestamped sibling) | `radixark/miles:latest-cu12` |
-| `rocm700-mi30x` / `rocm700-mi35x` / `rocm720-mi35x` | `rocm/sgl-dev:miles-rocm7xx-mi3xx` (+ timestamped sibling) | `rocm/sgl-dev:latest-rocm7xx-mi3xx` |
+| `rocm720-mi35x` / `rocm10-mi35x` | `rocm/sgl-dev:miles-rocm*-mi3xx` (+ timestamped sibling) | `rocm/sgl-dev:latest-rocm*-mi3xx` |
 
 What **moves a shared tag**: `--image-tag dev` overwrites `:dev` (or `:dev-cu12`) and adds a timestamped sibling; on a **scheduled** run `latest`→`dev` *and* `latest-cu12`→`dev-cu12` both advance; pruning likewise runs **only on schedule**, keeping the newest 20 of **each** series — `dev-<ts>` and `dev-cu12-<ts>` independently. Any `workflow_dispatch` — **including** `simulate_schedule` — writes its own tag(s) but never moves `latest` or prunes; only the real cron mutates published tags. See the trigger table above.
 
@@ -126,7 +125,7 @@ gh workflow run docker-build.yml -f variant=cu13-x86 -f image_tag=custom -f cust
 
 | input | required | values / default |
 | ----- | -------- | ---------------- |
-| `variant` | yes | `cu13` / `cu13-x86` / `cu13-aarch64` / `cu12-x86` / `rocm700-mi30x` / `rocm700-mi35x` / `rocm720-mi35x` |
+| `variant` | yes | `cu13` / `cu13-x86` / `cu13-aarch64` / `cu12-x86` / `rocm720-mi35x` / `rocm10-mi35x` |
 | `image_tag` | yes | `dev` / `latest` / `custom` |
 | `custom_tag` | no | tag name; required when `image_tag=custom` |
 | `dockerfile` | no | path to Dockerfile (default `docker/Dockerfile`) |

docs/developer/versions.md

@@ -92,7 +92,7 @@ fleet's image is.
 | `cu13` | `radixark/miles:dev` | `linux/amd64` + `linux/arm64`, one manifest. This is the daily image |
 | `cu13-x86` / `cu13-aarch64` | `radixark/miles:dev` | Single-arch rebuilds of the same image |
 | `cu12-x86` | `radixark/miles:dev-cu12` | `linux/amd64`, CUDA 12.9 legacy |
-| `rocm700-mi30x` / `rocm700-mi35x` / `rocm720-mi35x` | `rocm/sgl-dev:miles-rocm7xx-mi3xx` | Native |
+| `rocm720-mi35x` / `rocm10-mi35x` | `rocm/sgl-dev:miles-rocm*-mi3xx` | Native |
 
 `--image-tag dev` also publishes a timestamped sibling. Scheduled retention and manual tag behavior are documented in [Docker build](/ci/02-docker-build).