Skip to content

build: bump vllm-torchtpu to 0.33.0, which serves LoRA on TPU - #463

Merged
hodaaaaaaaaaa merged 1 commit into
openrl-dev:feat/tpufrom
hodaaaaaaaaaa:build/vllm-torchtpu-0.33.0
Nov 19, 2026
Merged

hodaaaaaaaaaa merged 1 commit into
openrl-dev:feat/tpufrom
hodaaaaaaaaaa:build/vllm-torchtpu-0.33.0

Conversation

@hodaaaaaaaaaa

@hodaaaaaaaaaa hodaaaaaaaaaa commented Nov 19, 2026 •

Copy link
Copy Markdown
Collaborator

Problem

The released vllm-torchtpu 0.32.0 serves full fine-tuning but not LoRA, so a TPU sampler image for LoRA jobs has to be built from vllm-torchtpu's main branch.

Change

The tpu-sampler extra pins vllm-torchtpu==0.33.0, which carries multi-LoRA, so one released image serves LoRA and full fine-tuning. The from-main build notes are removed from pyproject.toml, Dockerfile.tpu_sampler and docs/configuration.md.

0.33.0's TPUWorker now has its own reset_encoder_cache. vLLM fails at startup if a worker extension defines a method the worker already has, so the extension no longer defines it.

0.33.0 carries multi-LoRA, so the released sampler image serves LoRA and
full fine-tuning and the from-main build notes go away. Its TPUWorker also
defines reset_encoder_cache; vLLM refuses a worker extension that redefines
a worker attribute, so the extension drops its own copy.

Signed-off-by: Hoda Mohammadi <hodamo@google.com>
@hodaaaaaaaaaa
hodaaaaaaaaaa marked this pull request as ready for review November 19, 2026 18:45
@hodaaaaaaaaaa
hodaaaaaaaaaa merged commit 653e1c3 into openrl-dev:feat/tpu Nov 19, 2026
12 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants