[Misc]feat: adapt to vLLM main (0d4d334e...46a95815)

[Meihan-chen]

Main2Main Summary

Status: completed
Upstream range: 0d4d334eaa583b9c09aa4eb7538c22db99fd84b3..46a95815d344e43ebd6b4757632c55d03e90983a
Reached upstream commit: 46a95815d344e43ebd6b4757632c55d03e90983a
Steps: 8/8
CI suite: e2e-main2main

Result

The target vLLM commit 46a95815d344e43ebd6b4757632c55d03e90983a was fully reached. All 8 planned steps completed with CI verification, and vllm-ascend was successfully adapted to absorb 23 upstream commits. Only one minor adaptation was required (adding dtype=np.int32 to token_arange_np override in eagle_proposer.py for type consistency with upstream SpecDecodeBaseProposer). All other changes were inherited automatically through vLLM's module imports and class inheritance.

Completed Steps

Step	Upstream range	vllm-ascend commit	CI result	Summary
step-1	0d4d334e..27b85d2084c	4e7a6ac	passed	InternS2 Preview model support, thinking_token_budget bugfix, CPU/RISC-V kernels (not relevant to Ascend)
step-2	27b85d2084c..2676ab1e0b	3622a09	passed	Deprecation removal of old tokenizer/chat template import paths
step-3	2676ab1e0b..4b364f810e1	ae9bbbd	passed	LoRA layer OOT class support, SWA prefix-cache optimization (backward compatible)
step-4	4b364f810e1..75fd68c7a57	770cc65	passed	Pooling offline API refactoring into mixin class
step-5	75fd68c7a57..0fe755025467	b84c2bc	passed	Revision/code_revision propagation, DFlash FP8 fix, token_arange_np dtype adaptation
step-6	0fe755025467..e0a45f1455ce	8b3a966	passed	IPC weight sync optimizations for RL training
step-7	e0a45f1455ce..fb5bd03f51f986bf	0041731	passed	IR op priority moved to WorkerBase.init, MLA optimizations
step-8	fb5bd03f51f986bf..46a95815d34	8585eab	passed	ROCm MLA FP8 ASM prefill (not relevant to Ascend)

Changes Made

• Updated vLLM commit reference from 0d4d334eaa583b9c09aa4eb7538c22db99fd84b3 to 46a95815d344e43ebd6b4757632c55d03e90983a.
• Updated .github/workflows/dockerfiles/Dockerfile.lint, .github/workflows/pr_test_full.yaml, .github/workflows/pr_test_light.yaml, and docs/source/conf.py with new vLLM commit hash in each step.
• Minor adaptation: Added dtype=np.int32 to token_arange_np override in vllm_ascend/spec_decode/eagle_proposer.py (line 184) for type consistency with upstream SpecDecodeBaseProposer change.

CI Verification

• Passed: All 8 steps passed CI verification with ci_result: "passed" and run_suite_exit_code: 0.
• Treated as env flakes: none
• Last successful step: step-8

Key Upstream Changes Absorbed

1. InternS2 Preview model support - New model architecture registered and spec decode proposer list updated
2. Deprecation removals - Old get_tokenizer and resolve_hf_chat_template import paths removed
3. LoRA OOT support - Added maybe_get_oot_by_class() for out-of-tree layer support
4. SWA prefix-cache optimization - Skip caching SWA blocks that can never serve a prefix-cache hit
5. Pooling API refactoring - Split pooling offline API into PoolingOfflineMixin
6. Revision propagation - Pass revision/code_revision to all artifact boundaries in model loading
7. IPC weight sync - Added multigpu support and chunked packed tensors for RL training
8. IR op priority refactoring - Moved from set_forward_context() to permanent setup in WorkerBase.__init__
9. MLA optimizations - Various MLA attention performance improvements (ROCm-specific)

vllm-ascend Commits Created

All commits are signed and located in the vllm-ascend repository:

1. 4e7a6ac70b220ccb51d96099b35e70a19052c1d2 (step-1)
2. 3622a097ee6e7f225b7cc99cf0ee573c31ce1273 (step-2)
3. ae9bbbdf7f2a74ed0c3b8f54eb848b564d740372 (step-3)
4. 770cc65981c99d436cf3c78e3405593d7b1d881c (step-4)
5. b84c2bcbd389facd975711e546bdb47ad14be9b0 (step-5)
6. 8b3a966e467946d16d14f5ce066d50b113135f31 (step-6)
7. 0041731377b8a261939a439685ee6ca3462df084 (step-7)
8. 8585eabe70b8114192bcbd09ef8cdb02099311bc (step-8)

• vLLM version: v0.20.2
• vLLM main: vllm-project/vllm@0d4d334