Repository navigation
HRX: follow AMD's live ggml-hrx, pinned to AMD's tested pair and kept current - #11
Merged
Merged
Conversation
… current third_party/llama.cpp moves to the org fork 1bit-MONSTER/llama.cpp (of ggml-org/llama.cpp), branch 1bit/hrx-vulkan = AMD's hrx-graph-develop-v2 f1a0aca; third_party/hrx-system moves to fab1624. That is the pair ROCm/ggml-staging-automation builds and tests. The previous build sat on AMD's hrx-v2 line, abandoned in June and 2455 commits behind ggml-org; it is kept as 1bit/hrx2-archive in the fork, with our Q4NX kernels and zaya, for a later port to ggml-hrx. - cmake/hrx.cmake: one external project. llama.cpp's ggml-hrx builds hrx-system itself (HRX_SOURCE_DIR) with TheRock's amdclang. The separate ROCm/hrx runtime, its three patches and the loom-link-only build are gone. - The device is HRX0 (ggml-hrx), not HRX20 (ggml-hrx2): app defaults, the e2e test, docs. - HRX dlopens the HSA runtime; the distro's rejects gfx1151's PM4-emulation probe and HRX then registers no device. CMake finds TheRock's libhsa-runtime64 and `1bit lemonade` passes it to llama-server as IREE_HAL_AMDGPU_LIBHSA_PATH (unless already set). - .github/workflows/bump-hrx.yml: daily, when AMD moves its pair, syncs the fork and opens a PR moving both submodules (needs the HRX_BUMP_TOKEN secret). Strix Halo, Qwen3-0.6B: tests/hrx_lemonade_e2e.sh PASS (Paris on Vulkan0 and on HRX0). llama-bench pp512/tg128, previous -> now: Vulkan0 Q4_K_M 10627/306 -> 13905/338; HRX Q4_K_M 1046/68 -> 21702/303. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
bong-water-water-bong
enabled auto-merge (squash)
September 23, 2026 16:19
bong-water-water-bong
added a commit
that referenced
this pull request
Sep 26, 2026
…bit-MONSTER/llama.cpp #11) (#127) The pin adds sliding-window attention and a second rope base for ZAYA1-74B-preview on top of 1e775cd (#119). ZAYA1-8B is unchanged: Q4_K_M perplexity 21.5731 on Vulkan before and after, and test-llama-archs -a zaya passes. ZAYA1-74B-preview Q4_K_M passes serve_e2e through 1bit serve --device vulkan. registry/architectures.json, regenerated: it records the new pin (it still named 5556bf2 after #119 moved the pin to 1e775cd without regenerating it). Co-authored-by: bong-water-water-bong <bong-water-water-bong@1bit.gg> Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
bong-water-water-bong
pushed a commit
that referenced
this pull request
Sep 26, 2026
…oints and three HRX fixes Our llama.cpp #7 (HRX MoE router route-ids binding length), #9 (decode-split flash-attention capacity bound), #11 (ZAYA sliding-window attention, ZAYA1-74B), #12 (MoE router experts-per-wave launch geometry) and #14 (ZAYA legacy checkpoints; --remote fetches chat_template.jinja). 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.
Moves the engine's llama.cpp to AMD's maintained HRX line, pinned to the pair AMD tests, with a daily job that keeps it current.
Why
The engine pinned
bong-water-water-bong/llama.cpp@3b33c8a. That's AMD'shrx-v2line (ggml-hrx2, deviceHRX20) plus our 33 commits. AMD abandoned that line in June, and it's 2455 commits behind ggml-org. AMD's live line isAMD-Ecosystem/llama.cpphrx-graph-develop-v2, which has a singleggml-hrxbackend (deviceHRX0), was updated yesterday, and is 915 behind. It's pinned, together withROCm/hrx-system, by AMD's integration repoROCm/ggml-staging-automation.What
third_party/llama.cppnow points at the new org fork1bit-MONSTER/llama.cpp(fork ofggml-org/llama.cpp), branch1bit/hrx-vulkan= AMD'sf1a0aca.third_party/hrx-system=fab1624, AMD's pin. The separatethird_party/hrx(ROCm/hrx) submodule andpatches/hrx/are removed.cmake/hrx.cmake): a single external project. llama.cpp's ggml-hrx builds hrx-system itself (HRX_SOURCE_DIR) with TheRock's amdclang.libhsa-runtime64rejects gfx1151'sPM4_EMULATIONprobe, so HRX then registers no device. CMake now finds TheRock's copy, and1bit lemonadepasses it to llama-server asIREE_HAL_AMDGPU_LIBHSA_PATH(unless you've set it yourself).HRX20→HRX0in the app defaults,tests/hrx_lemonade_e2e.shand the docs..github/workflows/bump-hrx.yml. When AMD moves its pair, it syncs the fork (masterto ggml-org,1bit/hrx-vulkanto AMD's pin) and opens a PR moving both submodules. It needs a secretHRX_BUMP_TOKEN: a fine-grained token with Contents read/write on1bit-MONSTER/llama.cppand1bit-MONSTER/engine, and Pull requests read/write on1bit-MONSTER/engine.1bit/hrx2-archivein the fork (our Q4NX kernels, zaya, MoE grouping) for a later port to ggml-hrx.Verified on Strix Halo
tests/hrx_lemonade_e2e.shPASS. Qwen3-0.6B Q4_0 answers "Paris" through both recipes, served by this build's llama-server onVulkan0andHRX0.Lost until ported
Q4NX GGUFs on HRX: our Q4NX kernels lived in ggml-hrx2, which ggml-hrx doesn't have. GGUFs still route to Vulkan0.
🤖 Generated with Claude Code