Repository navigation
fix(ci-mesh): K3 AgentX cells — FP8 prefill attention and capture from bs=1 - #2417
Merged
Merged
Conversation
Sync the K3 P/D nightly cells with recipes/Agentic-Kimi-K3.md as of d8cdd4e: - ATOM_USE_FLYDSL_FP8_PREFILL_ATTN=1 on every band - PrefillDelayer (ATOM_PREFILL_DECODE_INTERVAL=4, ATOM_PREFILL_DELAYER_MAX_QUEUE_MS=5000) from CONC 16 up - `cudagraph: auto` captures the dense range from bs=1 instead of 2 Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Its benefit was measured on the aggregated recipe server; on the P/D split the decode instance never installs it and the prefill instance has no decode batch to protect, so keep the cells unchanged there. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
gbyu-amd
changed the title
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Brings two items from
recipes/Agentic-Kimi-K3.mdon main (#2382,
d8cdd4e2) into the Kimi-K3 P/D nightly cells. Targetszwan/feat-k3-pd-disagg-ci.ATOM_USE_FLYDSL_FP8_PREFILL_ATTN=1on every bandenv.common→ all 16 K3 cellscudagraph: auto→[2..graph_max][1..graph_max](pd_server_atom.sh; only K3 cells useauto)Deliberately not taken: the recipe's PrefillDelayer from CONC 16 up
(
ATOM_PREFILL_DECODE_INTERVAL=4,ATOM_PREFILL_DELAYER_MAX_QUEUE_MS=5000).Its gain was measured on the aggregated recipe server. It is unclear whether it helps on the P/D split: main
(#2238) never installs it on the decode instance (
disagg_is_decode), and theprefill instance has no decode batch to protect. This branch also predates
#2238, so it would currently only take effect in DP mode. The first commit added it and the second commit removes it,
so the net diff doesn't include it.
The rest of the recipe's per-concurrency table (DCP, spec/AL, ReplaySSM, LMCache 128/192 GiB,
max-num-seqs,AITER_REUSE_IDENTICAL_COMM_GROUPSfrom C56,state_checkpoint_interval_tokens=-1,ATOM_STATE_CHECKPOINT_DEMAND=0) already matched.Verification
pd_matrix.pybefore/after: across all 16 cells the only change isATOM_USE_FLYDSL_FP8_PREFILL_ATTN=1(64 = 16 cells × 4 role env maps).auto_cudagraph_capture_sizes: C1 spec7 →[1..16], C72 seqs144 →[1..144], matching the recipe'sgraph_max.tests/test_atomesh_pd_matrix.py,test_atomesh_spur_dispatch.py,test_atomesh_lmcache_disk.py: 26 passed (same as base).Notes
🤖 Generated with Claude Code