docker: upgrade base to vLLM 0.23.0, remove CUDA 13 build path

docker/Dockerfile

@@ -1,14 +1,11 @@
-ARG BASE_IMAGE=vllm/vllm-openai:v0.22.0-cu129-ubuntu2404
+ARG BASE_IMAGE=vllm/vllm-openai:v0.23.0-cu129-ubuntu2404
 FROM ${BASE_IMAGE}
 
 # ======================================== Arguments =============================================
 
 ARG PATCH_VERSION=latest
 ARG MEGATRON_COMMIT=1dcf0dafa884ad52ffb243625717a3471643e087
 
-ARG ENABLE_CUDA_13=0
-ARG TMS_CUDA_MAJOR=
-
 # ======================================== Setup =============================================
 
 WORKDIR /root/
@@ -53,33 +50,10 @@ RUN if [ "${INSTALL_FLASHQLA}" = "1" ]; then \
     fi
 RUN pip install tilelang -f https://tile-ai.github.io/whl/nightly/cu128/
 
-# cublas + (cu13) cuda dev headers. The arm64 (sbsa) vllm/vllm-openai base ships
-# the cublas runtime .so but not the dev header (cublas_v2.h) / unversioned .so
-# symlink that the x86 base has; TE needs the header and its CMake only creates
-# the CUDA::cublas target when both exist. The dev pkg must match the toolkit
-# CUDA major. For cu13 the top apt block's -12-9 dev set is the wrong major, so
-# also install the -13-0 nvrtc/nvtx/etc. headers TE's nvcc build needs. Placed
-# after the slow flash-attn layers so their cache is preserved; before TE.
-RUN apt-get update && \
-    if [ "${ENABLE_CUDA_13}" = "1" ]; then \
-      apt-get install -y libcublas-dev-13-0 \
-        cuda-nvrtc-dev-13-0 cuda-nvtx-13-0 cuda-nvml-dev-13-0 cuda-profiler-api-13-0 \
-        libcusparse-dev-13-0 libcusolver-dev-13-0 libcufft-dev-13-0 libcurand-dev-13-0; \
-    else apt-get install -y libcublas-dev-12-9; fi && \
-    rm -rf /var/lib/apt/lists/*
+# cublas dev header for TE CMake (arm64 base ships runtime .so but not the header).
+RUN apt-get update && apt-get install -y libcublas-dev-12-9 && rm -rf /var/lib/apt/lists/*
 
-# TE does not have wheel on cuda 13 yet, thus need to install from source.
-# TE is built with --no-build-isolation, so its build deps must be pre-installed;
-# install the set TE's release_v2.10 CI uses (wheel packaging ninja pybind11).
-# nvidia-mathdx is left unpinned (as in TE CI): 26.6.0 is x86-only, arm64 (sbsa)
-# max is 25.6.0, and TE release_v2.10 does not pin or reference it (it links plain
-# CUDA::cublas), so the resolver picking the per-arch latest is safe.
-RUN if [ "${ENABLE_CUDA_13}" = "1" ]; then \
-      pip install nvidia-mathdx pybind11 ninja wheel packaging && \
-      pip -v install --no-build-isolation git+https://github.com/NVIDIA/TransformerEngine.git@release_v2.10; \
-    else \
-      pip -v install --no-build-isolation "transformer_engine[pytorch]==2.10.0"; \
-    fi
+RUN pip -v install --no-build-isolation "transformer_engine[pytorch]==2.10.0"
 
 RUN NVCC_APPEND_FLAGS="--threads 4" \
   pip -v install --disable-pip-version-check --no-cache-dir \
@@ -89,20 +63,11 @@ RUN NVCC_APPEND_FLAGS="--threads 4" \
 RUN git clone https://github.com/NVIDIA/Megatron-LM.git --recursive && \
     cd Megatron-LM && git checkout ${MEGATRON_COMMIT}
 
-# torch_memory_saver pinned to a193d9dd (upstream slime #1916). The newer commit
-# ships a multi-CUDA wheel and requires TMS_CUDA_MAJOR at build time; default it
-# to the running torch's CUDA major (slime #1924).
-RUN TMS_CUDA_MAJOR="${TMS_CUDA_MAJOR:-$(python -c 'import torch; print(torch.version.cuda.split(".")[0])')}" && \
-    export TMS_CUDA_MAJOR && \
-    pip install git+https://github.com/fzyzcjy/torch_memory_saver.git@a193d9dd1b877d33c64a41cfb3db9f867df2d926 --no-cache-dir --force-reinstall
+# torch_memory_saver pinned to a193d9dd (upstream slime #1916).
+RUN pip install git+https://github.com/fzyzcjy/torch_memory_saver.git@a193d9dd1b877d33c64a41cfb3db9f867df2d926 --no-cache-dir --force-reinstall
 RUN pip install git+https://github.com/radixark/Megatron-Bridge.git@bridge --no-deps --no-build-isolation
 RUN pip install nvidia-modelopt[torch]>=0.37.0 --no-build-isolation
 
-# This patch from masahi will be included in later Triton releases
-RUN if [ "$ENABLE_CUDA_13" = "1" ]; then \
-    (cd /root && git clone -b feat/v350_plus_8045 https://github.com/fzyzcjy/triton.git && cd triton && pip install -r python/requirements.txt && pip install --verbose -e .); \
-  fi
-
 COPY requirements.txt /tmp/requirements.txt
 RUN pip install --ignore-installed PyJWT && \
     pip install -r /tmp/requirements.txt

docker/justfile

@@ -3,34 +3,23 @@
 # both the linux/amd64 and linux/arm64 digests and `docker pull` auto-selects
 # the platform. No arch suffix is ever published as a tag.
 #
-# CUDA 12.9 is the default, so it carries NO cu marker in the tag. Only the
-# non-default cu13 variant is suffixed.
-#
 # The CUDA + source-built FA/TE images can't be cross-emulated, so each arch is
 # built on its own native host and pushed BY DIGEST (no tag lands in the hub),
 # then the two digests are fused into the final tag with `just manifest`.
 #
 # Tag scheme:
-#   inferactinc/public:vime-<version>        immutable, multi-arch (cu12.9)
-#   inferactinc/public:vime-latest           rolling,   multi-arch (cu12.9)
-#   inferactinc/public:vime-cu13-<version>   immutable, multi-arch (cu13 variant)
-#   inferactinc/public:vime-cu13-latest      rolling,   multi-arch (cu13 variant)
+#   inferactinc/public:vime-<version>        immutable, multi-arch
+#   inferactinc/public:vime-latest           rolling,   multi-arch
 # <version> comes from docker/version.txt.
 
 IMAGE := "inferactinc/public"
 BUILDER := "vime-builder"
 
 # ---- per-arch build, pushed BY DIGEST (run once on an amd64 host, once on an arm64 host) ----
 
-# Default — cu12.9, no cu marker in the tag.
 build:
     ARG_TAG_SUFFIX="" ARG_BUILD_EXTRA_ARGS="--build-arg INSTALL_FLASHQLA=1" just _build-digest
 
-# cu13 variant — vLLM default-CUDA base; ENABLE_CUDA_13 builds the CUDA-13
-# TransformerEngine/Triton on top.
-build-cu13:
-    ARG_TAG_SUFFIX="-cu13" ARG_BUILD_EXTRA_ARGS='--build-arg BASE_IMAGE=vllm/vllm-openai:v0.22.0-ubuntu2404 --build-arg ENABLE_CUDA_13=1' just _build-digest
-
 _build-digest:
     #!/bin/bash
     set -euxo pipefail
@@ -48,19 +37,15 @@ _build-digest:
     jq -r '."containerimage.digest"' "$META"
 
 # ---- fuse the two per-arch digests into one multi-arch tag ----
-# Run once after `build` (or `build-cu13`) has pushed on BOTH hosts, passing the
-# digests it printed. For the default cu12.9 image leave VARIANT empty:
-#   just manifest "" sha256:<amd64> sha256:<arm64>     -> vime-<version> + vime-latest
-#   just manifest cu13 sha256:<amd64> sha256:<arm64>   -> vime-cu13-<version> + vime-cu13-latest
-manifest VARIANT AMD_DIGEST ARM_DIGEST:
+# Run once after `build` has pushed on BOTH hosts, passing the digests it printed:
+#   just manifest sha256:<amd64> sha256:<arm64>     -> vime-<version> + vime-latest
+manifest AMD_DIGEST ARM_DIGEST:
     #!/bin/bash
     set -euxo pipefail
     cd ..
 
     VERSION="$(cat docker/version.txt | tr -d '\n')"
-    SUFFIX=""
-    [ -n "{{VARIANT}}" ] && SUFFIX="-{{VARIANT}}"
-    docker buildx imagetools create -t "{{IMAGE}}:vime${SUFFIX}-${VERSION}" -t "{{IMAGE}}:vime${SUFFIX}-latest" "{{IMAGE}}@{{AMD_DIGEST}}" "{{IMAGE}}@{{ARM_DIGEST}}"
+    docker buildx imagetools create -t "{{IMAGE}}:vime-${VERSION}" -t "{{IMAGE}}:vime-latest" "{{IMAGE}}@{{AMD_DIGEST}}" "{{IMAGE}}@{{ARM_DIGEST}}"
 
 # ---- single-arch test/debug image for the run-ci-image validation job ----
 # The e2e-test-image runner is x86, so this is amd64-only and

docker/patch/latest/vllm.patch

@@ -1,30 +1,22 @@
-diff --git a/vllm/v1/engine/core.py b/vllm/v1/engine/core.py
---- a/vllm/v1/engine/core.py
-+++ b/vllm/v1/engine/core.py
-@@ -750,8 +750,10 @@
-         if tags is None or tags:
-             self.model_executor.wake_up(tags)
+diff --git a/vllm/model_executor/layers/fused_moe/all2all_utils.py b/vllm/model_executor/layers/fused_moe/all2all_utils.py
+--- a/vllm/model_executor/layers/fused_moe/all2all_utils.py
++++ b/vllm/model_executor/layers/fused_moe/all2all_utils.py
+@@ -5,7 +5,6 @@ from typing import Any
 
--        # Resume scheduling (applies to all levels)
--        self.resume_scheduler()
-+        # Partial wakes intentionally keep the remaining allocations asleep.
-+        # Resume scheduling only once all executor memory is resident again.
-+        if not self.model_executor.is_sleeping:
-+            self.resume_scheduler()
+ import torch
 
-     def is_sleeping(self) -> bool:
-         """Check if engine is sleeping at any level."""
-@@ -1844,8 +1846,11 @@
-                     continue
+-from vllm.config import get_current_vllm_config
+ from vllm.distributed import (
+     get_ep_group,
+ )
+@@ -240,9 +239,7 @@ def maybe_make_prepare_finalize(
 
-                 # We are in a running state and so must execute a dummy pass
--                # if the model didn't execute any ready requests.
--                self.execute_dummy_batch()
-+                # if the model didn't execute any ready requests -- unless the executor is
-+                # asleep (#44483: a decode-shaped dummy batch reads freed KV -> illegal memory
-+                # access). The finished-sync all-reduce below still runs (DP lockstep).
-+                if not self.is_sleeping():
-+                    self.execute_dummy_batch()
-
-             # 3) All-reduce operation to determine global unfinished reqs.
-             self.engines_running = self._has_global_unfinished_reqs(
+     elif moe.use_fi_nvl_one_sided_kernels:
+         assert quant_config is not None
+-        max_num_tokens = (
+-            get_current_vllm_config().scheduler_config.max_num_batched_tokens
+-        )
++        max_num_tokens = moe.max_num_tokens
+         if quant_config.quant_dtype is None:
+             dispatch_dtype_bytes_per_elem = 2
+             dispatch_scale_bytes_per_token = 0