Skip to content

[recipe] add Agentic-K3 recipe - #1941

Open
gbyu-amd wants to merge 2 commits into
mainfrom
guanbao/update-k3-agentic-recipe
Open

[recipe] add Agentic-K3 recipe#1941
gbyu-amd wants to merge 2 commits into
mainfrom
guanbao/update-k3-agentic-recipe

Conversation

@gbyu-amd

Copy link
Copy Markdown
Contributor

Motivation

Technical Details

Test Plan

Test Result

Submission Checklist

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
@github-actions

Copy link
Copy Markdown
Contributor

🏷️ 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 1941 --add-label <label>

@zufayu
zufayu requested a review from valarLip August 19, 2026 01:24
con8 ran with neither tier: 32 seqs, 96 checkpoint slots, all of it in
HBM. Moving it onto the same LMCache + state-offload pair con10 uses --
16 seqs, 16 slots, 32 GiB/rank of CPU KV -- measures faster, so that is
what the profile should say. ReplaySSM stays on and the batched-token
budget stays at 8192, which is where con8 still differs from con10.

Three defaults the shared blocks got wrong for anyone who enables them
outside a named profile, all found the hard way:

  - LMCACHE_MAX_LOCAL_CPU_SIZE fell back to 200 GiB. Each rank allocates
    its own, and 200 hangs the launch mid-pin (seen at rank 2). 32.
  - OFFLOAD_GPU_STAGING_CHUNKS was set but unexplained. The default is
    2 chunks (8 MiB), one K3 state entry is 54.78 MiB, and a buffer that
    cannot hold one entry makes the tier decline to build -- one log
    line, then nothing offloads, which reads exactly like a tier that is
    on and idle.
  - OFFLOAD_SAVE_PER_REQ_CACHE has no line here and should not get one.
    A sequence owning per-request state has its KV load leg refused, so
    saving that KV costs ~8% throughput for bytes nothing reads back.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
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