chore: relaunch vllm 0.20.1 bump

[samsja]

Re-applies #2427 (vllm 0.20.1 bump) after revert in #2437.

Why

• The original bump (bump vllm version #2427) was reverted in Revert "bump vllm version (#2427)" #2437 without a recorded blocker. The intervening fp32 lm_head feature (feat(inference): fp32 lm_head via native bf16xbf16 -> fp32 mm (alt to #2438) #2441) actually relies on torch.mm(..., out_dtype=torch.float32), which needs torch >= 2.10 — comfortably available with the new flash-attn wheel pinned to torch 2.11 here.
• Verified the vLLM LogitsProcessor API is byte-identical between 0.19.0 and 0.20.1 (__init__, _get_logits, _gather_logits signatures unchanged), so monkey_patch_fp32_lm_head from feat(inference): fp32 lm_head via native bf16xbf16 -> fp32 mm (alt to #2438) #2441 keeps working without modification.

What

• Reverts commit bb522cc ("Revert 'bump vllm version (bump vllm version #2427)'"), which restores:
  • vllm>=0.20.1
  • flash-attn wheel pinned to +cu128torch2.11
  • torchvision / torchaudio deps via pytorch-cu128
  • removal of monkey_patch_fused_moe_lora_dp (fixed in vLLM 0.20 via #40338)
  • removal of monkey_patch_offloading_connector_cpu_block_count (fixed in vLLM 0.20 via #39617)
• monkey_patch_fp32_lm_head (feat(inference): fp32 lm_head via native bf16xbf16 -> fp32 mm (alt to #2438) #2441) is preserved as-is — auto-merge handled the overlap with the patches removed by the bump.

🤖 Generated with Claude Code

---

Note

Medium Risk
Upgrades core inference dependencies (vLLM/PyTorch/Flash-Attn), which can change runtime behavior and performance across serving and distributed execution despite minimal application-code changes.

Overview
Updates the inference dependency stack to vLLM >=0.20.1 and aligns CUDA 12.8 wheels by adding torchvision/torchaudio from the pytorch-cu128 index and updating the pinned flash-attn wheel to the +cu128torch2.11 build.

Simplifies local vLLM patching by removing monkey patches that are now fixed upstream (the LoRA+MoE+DP corruption workaround and the offloading connector CPU block count fix), and adjusts comments/config (exclude-newer-package now unblocks vllm).

^{Reviewed by Cursor Bugbot for commit 464bea4. Bugbot is set up for automated code reviews on this repo. Configure here.}