feat(vllm): compile GLM-5.3 native support on 0.28

[sihang-s]

Summary

• build only the compile-sensitive GLM-5.3-Flash subset into the vLLM 0.28 wheel
• pin FlashInfer to the immutable source for nightly-v0.6.18-20260819 while keeping CUDA 13.2.1
• remove the obsolete local FlashInfer routing patch now included upstream
• preserve the existing CoreWeave MNNVL native patch
• leave the Python/runtime files and the two-file narrow block-table fix for downstream runtime layering

Compiled subset

• cmake/external_projects/vllm_flash_attn.cmake
• csrc/libtorch_stable/cache_kernels.cu

These files must exist before wheel/native-extension compilation because they pin the flash-attention native dependency and add compiled head_dim=512 cache-kernel support.

Source provenance

• CoreWeave base: vLLM 2cf0a6915ce544dc493a0990f2ea38d81601128a
• Upstream PR: [Model] add GLM-5.3-Flash support vllm-project/vllm#53906
• Reviewed upstream head: 0f82208cb1d01e723bb25b607217a1adfb0cfc30
• Original feature commit: 933876c388fb129ad82590660e6506614559cb86
• FlashInfer: ee3fda10e0b9d82b71bda534bc2954d3176e6ded (nightly-v0.6.18-20260819)
• FlashInfer no-group routing fix: 08ddfbcd2e89b2f4b68391825817909e30d445e2 (already included in the pinned source)
• flash-attention: vllm-project/flash-attention@4eb29359a51b4fe67c8c0f0c5e931236971aa492
• Compile patch SHA256: 00e84751bd79e3140fc4344223d9997548605d47ba66f4ee51d108520831d762

Existing native patch consolidation

• moved the MNNVL Lamport mailbox fix into the numbered vllm-patches/ stack
• refreshed it to the final upstream source hunk from vLLM #53000 at a047e2543da570a64d1bbfeac4fe44eff3e87a81
• retained it because the pinned v0.28.0 source predates the fix; the first containing release tag is v0.28.1rc0
• MNNVL patch SHA256: d4e98e4b302ecdaf463120b7ca1b1f567c18525b742ae83d2d9a445c042173e9

Compatibility contract

The downstream runtime patch must use the same vLLM base and reviewed upstream head. Python/runtime-only upstream deltas update only the downstream layer; CMake, CUDA/C++, native-binding, native-dependency, or FlashInfer changes require a new compiled image and digest.

Validation

• compile-only patch applies cleanly to the exact vLLM 0.28 source pin
• existing MNNVL patch applies after the compiled subset
• Docker buildx checks pass for linux/amd64 and linux/arm64 with no warnings
• vllm-downloader and flashinfer-downloader stages pass for linux/arm64
• full amd64/arm64 wheel and image builds are delegated to this PR's CI

Follow-up

After CI publishes the new immutable multi-arch digest, downstream consumers can pin this compiled image and layer the Python/runtime GLM backport.

[nehashah-eng]

@alexeldeib also has other patch here - https://github.com/coreweave/ml-containers/tree/main/vllm-tensorizer ... maybe we should use all the patches in the same way.

[alexeldeib]

seems okay assuming it works 🤞

Let's strip out the 2 forks as soon as convenient in follow up when upstream is ready

[github-actions]

@sihang-s Build complete, success: https://github.com/coreweave/ml-containers/actions/runs/33115583554
Image: ghcr.io/coreweave/ml-containers/vllm-tensorizer:codex-vllm-028-glm53-7485122-2cf0a6915ce544dc493a0990f2ea38d81601128a-cuda13.2.1-ubuntu24.04

[github-actions]

@sihang-s Build complete, success: https://github.com/coreweave/ml-containers/actions/runs/33140149775
Image: ghcr.io/coreweave/ml-containers/vllm-tensorizer:codex-vllm-028-glm53-8292f08-2cf0a6915ce544dc493a0990f2ea38d81601128a-cuda13.2.1-ubuntu24.04

[github-actions]

@sihang-s Build complete, success: https://github.com/coreweave/ml-containers/actions/runs/33155124037
Image: ghcr.io/coreweave/ml-containers/vllm-tensorizer:codex-vllm-028-glm53-bcbe5f0-2cf0a6915ce544dc493a0990f2ea38d81601128a-cuda13.2.1-ubuntu24.04

[nehashah-eng]

Just to confirm this is a release right, not an arbitrary commit ?

[sihang-s]

It's from a RC tag. Pinging commit SHA here for immutable build.

[sihang-s]

v0.6.18rc10