Repository navigation
docs/hrx.md: known issue — --device hrx answers wrongly on mid-length prompts - #45
Merged
Merged
Conversation
…oken prompts (the prefill split is fine) Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… 256 tokens (prefill is right) 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
… on HRX (#239) llama.cpp #45: the K-quant SwiGLU decode kernel wins over AMD's lossy int4 lowrow path (27B llama-server decode 10.7 -> 12.0 tok/s; KLD vs CPU 0.00167, AMD kernels 0.00318). llama.cpp #46: 2-8 token K-quant kernels for speculative verify: Qwen3.8-27B + MTP head 20.1/16.4/17.6 -> 22.5/17.7/19.0 tok/s (Vulkan 34.6/24.3/24.5). docs/hrx.md updated. Co-authored-by: bong-water-water-bong <bong-water-water-bong@1bit.gg> Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
This was referenced Oct 7, 2026
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
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Found while checking HRX0 for the fused-kernel race seen on the hrx-v2 line: on our pinned pair,
--device hrxgarbles answers for prompts of roughly 404-1,733 tokens (Qwen3-0.6B, greedy, llama-server), while Vulkan is right at every length and HRX0 is right at 146 and >= 2,167 tokens. Context size and every HRX diagnostic switch leave it unchanged.Also recorded: Qwen3.6-35B-A3B Q8_0 faults the GPU on HRX0, and multi-sequence batches stop on a claimed-but-unrunnable FLASH_ATTN_EXT.
The prefill split is not affected:
--device vulkan --prefill-device hrxmatches Vulkan alone at 1,160 / 1,733 / 2,844 tokens (HRX0 prefills whole ubatches only; Vulkan decodes).This PR only documents it, so nobody serves on
--device hrxmeanwhile. The fix belongs in ggml-hrx upstream; per the AMD-repo rule, the report is yours to write.🤖 Generated with Claude Code