Repository navigation
serve: one slot for Qwen3.5/3.8 on HRX when --parallel is not given - #244
Merged
Merged
Conversation
Without --parallel, serve passes no -np and llama-server opens its default several slots. On HRX0 a second concurrent request then shares a batch with the first, and Qwen3.5/3.8 (qwen35, qwen35moe, qwen3next; Ternary Bonsai too) fail with HTTP 500 at the multi-sequence softplus / GATED_DELTA_NET (found by the leaderboard session: Bonsai GSM8K died on its second request). serve now passes -np 1 for those architectures on HRX; requests queue instead. Complements #243 (-np 1 for an explicit --parallel 1) and #240 (--parallel >1 refused for them on HRX). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
bong-water-water-bong
added a commit
that referenced
this pull request
Sep 30, 2026
… route, pin llama.cpp b8d587e (#248) * Ternary Bonsai on HRX: exact PTQ1_0 -> Q4_0 converter and the prism.hadamard route PrismML's Ternary-Bonsai-2-27B (Qwen3.8-27B trained ternary) now runs on HRX0. - tools/ternary_to_q4_0.py writes PrismML's PTQ1_0 weights as Q4_0, bit for bit (q = trit + 8, the group's fp16 scale in each of its four blocks), decodes every group back before keeping the file, keeps all metadata and stamps onebit.ternary_q4_0 = 128. tests/ternary_to_q4_0_test.py checks it against PrismML's own encoder (run in CI next to the Q4NX repack test). - 1bit serve sends a prism.hadamard file to --device hrx (auto picks it; other devices are refused, upstream llama.cpp would ignore the rotation and answer garbage) and refuses PrismML's own ternary types with the converter's name. tests/prism_route.sh. - docs/hrx.md: how to run it, what was checked, and the one-sequence limit of the Qwen3.5 / Qwen3.8 gated delta net on HRX0. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * Pin llama.cpp b8d587e: load Hadamard-folded weights (llama.cpp #34) acf9c74 -> b8d587e is llama.cpp #34 only: src/llama-hadamard.{h,cpp} reads prism.hadamard.* and rotates the activations before each folded weight, so Ternary Bonsai runs on HRX0. docs/hrx.md: the Qwen3.5 / Qwen3.8 one-sequence note now points at serve's one-slot default and --parallel refusal (#244). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * registry: regenerate for the llama.cpp b8d587e pin Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> --------- Co-authored-by: bong-water-water-bong <bong-water-water-bong@1bit.gg> Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
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.
Fixes HTTP 500s on
1bit serve --device hrxfor Qwen3.5/3.8 models (qwen35, qwen35moe, qwen3next, including Ternary Bonsai) when two requests overlap and--parallelisn't given.The failure: without
--parallel, serve passes no-np, so llama-server opens its default several slots. A second concurrent request then shares a batch with the first. That hits the multi-sequence softplus /GATED_DELTA_NETpath, which has no HRX kernel yet, and fails with HTTP 500. The leaderboard session found it: its Bonsai GSM8K run died on its second request.The fix: for those architectures on HRX with no
--parallel, serve now passes-np 1. Requests queue instead of failing.Related PRs:
-np 1for an explicit--parallel 1. This PR changes a different part ofserve.cpp, so the two don't conflict.--parallelabove 1 for these models on HRX.Verification:
serve.cpppasses a syntax-only compile against the vendored headers. I didn't run it end-to-end; CI builds it.🤖 Generated with Claude Code