[feat](vllm): upgrade vllm to 0.26.1

[zejunchen-zejun]

Motivation

Upgrade the ATOM vLLM out-of-tree plugin from the existing vLLM 0.26.0 baseline (568afb3a1) to vLLM 0.26.1 (adbf08d97) while preserving model accuracy, speculative decoding, CUDA Graph execution, and hybrid recurrent-state behavior.

vLLM 0.26.1 changes CacheConfig and SchedulerConfig fields, KV-cache binding, attention backend capabilities, GDN/KDA metadata handling, block reuse, speculative decoding, and model post-load hooks. This PR adapts ATOM to these contracts without replacing the ROCm PyTorch, Triton, AITER, or FlyDSL stack supplied by the base image.

Technical Details

• Updated the pinned vLLM revision to:
  • Commit: adbf08d977fb3fa26c4f19826745a02abd6dd7ca
  • Version: 0.26.1rc1.dev305+gadbf08d97
• Preserved the base ROCm runtime:
  • PyTorch 2.10.0+rocm7.2.4
  • Triton 3.7.0
  • Transformers 5.12.1
• Continued building and installing the vLLM wheel without pulling the upstream PyTorch 2.13 build dependency.
• Kept AITER/FlyDSL ownership in the base image or CI overlay; the release Dockerfile does not pin or replace AITER.

Plugin compatibility

• Replaced removed CacheConfig.calculate_kv_scales accesses with capability-safe fallbacks.
• Updated ATOM attention layers for vLLM's bind_kv_cache() contract.
• Added compatibility for the removed SchedulerConfig.max_num_partial_prefills.
• Updated MHA/MLA backend capability reporting, kernel block-size selection, sliding-window support, and PCP rejection.
• Updated sparse-MLA metadata to use vLLM 0.26 host-resident sequence-length data.
• Updated speculative decoding and EAGLE3 compatibility patches for the new model-runner contracts.
• Preserved current PyTorch 2.10 stable-ABI integration.

DeepSeek-V4

• Identified vLLM commit a82f1b38 as the source of the DeepSeek-V4 block-reuse accuracy regression.
• Restored non-immediate block reuse only for ATOM stateful proxy/Mamba cache pools.
• Kept the upstream locality optimization for ordinary MHA/MLA models.
• Preserved the latest quant_mode mapping from main while resolving the rebase conflict.

Kimi-K3

• Added compatibility with vLLM's process_weights_after_loading(act_dtype) call.
• Switched Kimi-K3 from generic GDN metadata to the dedicated KimiK3KDAMetadata contract.
• Used the KDA-specific state shape, dtype, and state-copy helpers.
• Added a KDA-only FULL-CUDA-Graph adapter that compacts padded request rows and binds the correct recurrent-state slots.
• Kept generic GDN behavior unchanged for Qwen3.5 and other GDN models.
• Restored client64/server-active64 accuracy without the previous max-num-seqs=32 workaround.

Documentation

• Added the vLLM 0.26.1 upgrade report and compatibility tests.
• Added Kimi-K3 TP8 launch and accuracy guidance.
• Added direct EvalScope GPQA-Diamond commands for:
  • 26-question smoke testing
  • Full one-repeat upgrade validation
  • Official recommended eight-repeat evaluation
• Configured GPQA with max reasoning effort, streaming, temperature 1.0, top-p 0.95, and a 96K output budget.

Test Plan

Test Environment

• GPU: MI355 / gfx950
• Image: localhost/atom-vllm:v0.26.1
• vLLM: adbf08d977fb3fa26c4f19826745a02abd6dd7ca
• PyTorch: 2.10.0+rocm7.2.4
• Triton: 3.7.0
• Transformers: 5.12.1
• AITER validation wheel: 0.1.1.dev1+gaf02117fb
• FlyDSL: 0.3.0
• ATOM branch: vllm_026

Plugin compatibility tests:

pytest \
  tests/plugin/test_vllm_kimi_k3.py \
  tests/plugin/test_vllm_026_compat.py \
  tests/plugin/test_vllm_mha_backend.py \
  -q

[github-actions]

🏷️ CI Guide

Runs automatically on every eligible PR before approval:

• ✅ Pre Checkin: Black, Ruff, catalog schema validation, non-GPU unit tests

Heavy model tests:

• ✅ Run after the PR is approved and Pre Checkin passes
• ✅ Run immediately when an approval review is submitted
• ✅ Can be requested before approval with labels

Label	Tests
ci:full	Run all heavy PR model tests: native ATOM, vLLM, and SGLang
ci:atom	Run native ATOM model accuracy tests
ci:vllm	Run ATOM vLLM OOT model accuracy tests
ci:sglang	Run ATOM SGLang model accuracy tests

Heavy jobs are skipped when the PR is not approved and no matching ci:* label is present.
Add labels via the sidebar or gh pr edit 1794 --add-label <label>

[github-actions]

⚠️ [ruff] <I001> _{reported by reviewdog 🐶}
Import block is un-sorted or un-formatted

ATOM/atom/plugin/vllm/gdn_backend.py

Lines 8 to 19 in f86ab0a

	from __future__ import annotations
	import logging
	from vllm.v1.attention.backends.gdn_attn import (
	GDNAttentionBackend,
	GDNAttentionMetadataBuilder,
	)
	from vllm.v1.attention.backends.registry import (
	MambaAttentionBackendEnum,
	register_backend,
	)

[github-actions]

⚠️ [ruff] <I001> _{reported by reviewdog 🐶}
Import block is un-sorted or un-formatted

ATOM/atom/plugin/vllm/models/kimi_k25.py

Lines 1 to 33 in f86ab0a

	from typing import Any, Iterable, Optional, Union
	import torch
	from aiter.dist.parallel_state import get_pp_group
	from torch import nn
	from vllm.model_executor.models.kimi_k25 import (
	KimiK25DummyInputsBuilder,
	KimiK25ForConditionalGeneration as vLLMKimiK25,
	KimiK25MultiModalProcessor,
	KimiK25ProcessingInfo,
	)
	from vllm.model_executor.models.kimi_k25_vit import (
	KimiK25MultiModalProjector,
	MoonViT3dPretrainedModel,
	)
	from vllm.model_executor.models.vision import is_vit_use_data_parallel
	from vllm.multimodal import MULTIMODAL_REGISTRY
	from atom.config import Config, QuantizationConfig
	from atom.model_config.kimi_k25 import KimiK25Config
	from atom.model_loader.loader import WeightsMapper, load_model_in_plugin_mode
	from atom.model_ops.embed_head import ParallelLMHead, VocabParallelEmbedding
	from atom.model_ops.layernorm import RMSNorm
	from atom.models.deepseek_v2 import DeepseekV2DecoderLayer, DeepseekV2Model
	from atom.models.utils import (
	IntermediateTensors,
	PPMissingLayer,
	make_empty_intermediate_tensors_factory,
	make_layers,
	maybe_prefix,
	)
	from atom.plugin.vllm.model_wrapper import ATOMForConditionalGeneration
	from atom.utils.decorators import support_torch_compile