Repository navigation
Laya: long documents go to HRX (no ROCm in the route policy) - #276
Merged
Merged
Conversation
The engine's GPU route is HRX only since 2026-10-01. Long documents went to the lean ROCm build while HRX prefilled Qwen3.8-27B at 97-99 tok/s; since the HRX prompt-matmul routing (fork #55, engine #273) it prefills 335 tok/s pp512 and a 14,435-token prompt through serve at 265 tok/s. Every Laya class now routes to HRX. Past 32K of context HRX needs fork #59 (pinned separately). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
Docs7 for 1bit-monster/engine
Commit |
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
1 participant
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Moves Laya's
long_docclass from the lean ROCm build to HRX. This is the owner's 2026-10-01 direction: HRX only, zero ROCm.Why it's safe now. Long documents went to ROCm while HRX prefilled Qwen3.8-27B UD-Q4_K_XL at 97-99 tok/s. Since the HRX prompt-matmul routing (fork #55, pinned by #273):
1bit serve --device hrxat 265 tok/s, same text (bcloud-ae).Fork #61, in review, lifts prompt remainders further (pp1862 172 -> 235).
Hold until fork #59 is pinned. HRX failed prompts past 32,768 KV rows (
copy_transpose_f16's range), and #59 fixes that (pp512 at d32768: fails -> 227 tok/s). Long documents are where that matters.servedefaults HRX to a 32K context, so the default path is fine today, but this should merge after the #59 pin.Changes:
config/route-policy.json:long_doc->hrx, with the measured reason; the about text says what HRX doesn't run, Lemonade serves.docs/laya.mdanddocs/serve.md: describe the HRX-only policy.The Laya route tests use their own policy file (
tests/route-policy-e2e.json), so they're unaffected.🤖 Generated with Claude Code