mudler / mudler/vllm.cpp

Tenstorrent host-free decode graph: capture, replay, and on-device state advance

Open
#1,105 0 comments 0 reactions 0 assignees View on GitHub

Nobody has claimed this yet.

Dominant language
C++
Stars
423
Forks
53
Avg merge
20h 26m
Merged PRs (30d)
310

Description

Row: BACKEND-TENSTORRENT-HOST-FREE-FORWARD

Summary

The Tenstorrent Blackhole decode path cannot run a captured mesh-trace while any to_vector / host write sits inside the captured region. The trace-runner spike measured that flipping the T=1 hybrid thresholds is not enough: capture aborts on Reads are not supported during trace capture.

This issue owns:

  • BACKEND-TENSTORRENT-TRACE-RUNNER (SPIKE, decision record)
  • BACKEND-TENSTORRENT-HOST-FREE-FORWARD (ACTIVE: R1-R3b plus R2 on-device cur_pos advance)

What landed on PR #1053

Env-gated VT_TT_HOST_FREE_DECODE decode-graph capture. On a Blackhole P150, Qwen3-0.6B "Hello" at 80 tokens completed 79 replays with no hang, 5.8x vs eager, 22/22 argmax vs the per-step-copy baseline. Default path is inert.

Still owed

  • Fresh review of the host-free decode graph
  • Operator rerun of the 80-token no-hang gate and the TT golden on card
  • test_qwen3_paged_engine still times out under the flag

Specs

  • .agents/specs/tenstorrent-host-free-forward.md
  • .agents/specs/tenstorrent-host-free-r1.md
  • .agents/specs/tenstorrent-host-free-r2.md
  • .agents/specs/tenstorrent-trace-runner.md

Contributor guide

Open the contributing guide

First steps

  1. Read the whole issue, then the project's contributing guide.
  2. Comment on the issue to say you are picking it up — it saves two people doing the same work.
  3. Fork the repository and make your change on a branch.
  4. Open a pull request that references the issue number.

Research direction

Start with .agents/specs/tenstorrent-host-free-forward.md, the R1/R2 specs, and PR #1053; inspect the captured decode graph and the test_qwen3_paged_engine timeout. Done means completing the fresh review, rerunning the 80-token no-hang gate and TT golden on card, and documenting the timeout outcome.

Written by the indexing model from the issue text.

Assessment

Tech stack
cpp
Domain
ai-infra-agents, backend
Issue type
Feature
Difficulty
4/5
Estimated time
3-5 days
Activity status
Quiet
Clarity
Mostly clear
Newbie friendliness
35/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.