Step 2: HRX + Vulkan in one llama.cpp build, wired into Lemonade - #7
Merged
Merged
Conversation
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- Pinned submodules: ROCm/hrx 0bc22fb (runtime), ROCm/hrx-system 6743075f (loom-link with --mode=selective), and the llama.cpp fork on branch 1bit-engine/hrx-vulkan (3b33c8a9). patches/hrx 0001-0003 hold the local changes: build plumbing plus the gfx1151 PM4 probe fix. - cmake/hrx.cmake (ONEBIT_HRX, off by default): clones each submodule into the build tree, patches it there, and builds only what ggml-hrx2 needs. hrx_assemble_prefix.cmake reproduces the previously hand-assembled prefix. - 1bit lemonade uses the resulting llama-server for llamacpp-hrx (HRX20) and for llamacpp's Vulkan backend (Vulkan0), replacing only defaults. - Lemonade delta: an hrx_device option (default HRX0) replaces the hard-coded device. - Verified on Strix Halo from a fresh clone (217 s): 48/48 kernel artifacts byte-identical to the working build; decode parity; Lemonade serves the same checkpoint on Vulkan0 (117 tok/s) and HRX20 (34 tok/s) (tests/hrx_lemonade_e2e.sh). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
bong-water-water-bong
force-pushed
the
hrx-vulkan
branch
from
September 22, 2026 21:42
60b1dd5 to
2dc363b
Compare
bong-water-water-bong
force-pushed
the
embed-lemonade
branch
from
September 22, 2026 21:42
39d6152 to
b07eb96
Compare
bong-water-water-bong
added a commit
that referenced
this pull request
Sep 22, 2026
…#6) * Step 1: embed Lemonade's server core in the 1bit binary - third_party/lemonade: Lemonade v11.9.0, vendored complete, plus the local deltas in its UPSTREAM.md (the onebit backend, /v1/registry, the hrx-b66 pin, the embed CMake patch). Copied from 1bit-MONSTER main (last changed 256e68dd7). - app/main.cpp: `1bit lemonade [lemond options]`, ported from run_embedded_lemonade in tools/unified_server.cpp. FLM pinning is dropped (FLM is eliminated); the engine-registry injection returns with the NPU engine (step 3). - tests/smoke_lemonade.sh (ctest + CI): health ok, catalog with llamacpp-hrx, system-info. Uses scratch dirs. - docs/lemonade.md; porting map adds step 5 (every HF model: registry + daily census). third_party/ is exempt from our copyright notice. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * ci: install libdrm-dev (Lemonade's server core includes libdrm/drm.h) Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * smoke: require HRX catalog models only where Lemonade reports HRX supported CI runners have no AMD GPU, so Lemonade correctly hides llamacpp-hrx there. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * Step 2: HRX + Vulkan in one llama.cpp build, wired into Lemonade (#7) * wip: step 2 HRX + Vulkan build (testing from a fresh clone) Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * wip: pin the HRX submodules (gitlinks) Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * Step 2: HRX + Vulkan in one llama.cpp build, wired into Lemonade - Pinned submodules: ROCm/hrx 0bc22fb (runtime), ROCm/hrx-system 6743075f (loom-link with --mode=selective), and the llama.cpp fork on branch 1bit-engine/hrx-vulkan (3b33c8a9). patches/hrx 0001-0003 hold the local changes: build plumbing plus the gfx1151 PM4 probe fix. - cmake/hrx.cmake (ONEBIT_HRX, off by default): clones each submodule into the build tree, patches it there, and builds only what ggml-hrx2 needs. hrx_assemble_prefix.cmake reproduces the previously hand-assembled prefix. - 1bit lemonade uses the resulting llama-server for llamacpp-hrx (HRX20) and for llamacpp's Vulkan backend (Vulkan0), replacing only defaults. - Lemonade delta: an hrx_device option (default HRX0) replaces the hard-coded device. - Verified on Strix Halo from a fresh clone (217 s): 48/48 kernel artifacts byte-identical to the working build; decode parity; Lemonade serves the same checkpoint on Vulkan0 (117 tok/s) and HRX20 (34 tok/s) (tests/hrx_lemonade_e2e.sh). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
This was referenced Sep 25, 2026
Closed
bong-water-water-bong
added a commit
that referenced
this pull request
Sep 26, 2026
…es (#119) Was 4 commits behind 1bit/hrx-vulkan-patched, missing every landed fix for both open HRX issues: - engine#108 (dense on HRX0, MoE experts on another device): route-ids binding-length fix (llama.cpp#7) plus the fused router launch-geometry fix (llama.cpp#12) that PR #7's description claimed was included but wasn't - verified separately this session against the exact repro (-dev HRX0,Vulkan0 -ts 1,0 -ot exps=Vulkan0, Qwen3-Coder-30B-A3B): correct and deterministic across repeated prompts, disabling the fused dispatch reproduces the issue's documented fallback error exactly. - engine#115 (HRX flash-attention decode-split all_rejected above 2048 KV tokens): capacity-cap fix (llama.cpp#9) - falls through to the general wmma kernel above the cap instead of crashing. Built onebit (-DONEBIT_HRX=ON) against the new pin and re-verified both fixes end to end before this bump. Co-authored-by: agent <agent@local> Co-authored-by: Claude Sonnet 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.
Step 2 of the port (
docs/PORTING.md). Stacked on #6, which is stacked on #5.What
ROCm/hrx0bc22fb: the runtime.ROCm/hrx-system6743075f: the lastloom-linkwith--mode=selective, which the fork's kernel catalog needs.1bit-engine/hrx-vulkan(3b33c8a9).patches/hrx/0001–0003replace edits that until now existed only in a working tree. 0001 and 0002 are build plumbing. 0003 is a real runtime fix: the gfx1151 PM4 probe.-DONEBIT_HRX=ON(off by default) builds all three from the build tree, leaving the submodules clean.hrx_assemble_prefix.cmakereproduces the prefix that used to be hand-assembled and undocumented.1bit lemonaderunsllamacpp-hrxon HRX20 andllamacpp/Vulkan on Vulkan0, both from this onellama-server. User-set config keys win.hrx_deviceoption (defaultHRX0) replaces the hard-coded device name. It's documented inUPSTREAM.md.Verified on Strix Halo (fresh clone, submodules at pins, 217 s build)
tests/hrx_lemonade_e2e.sh): the same checkpoint via both recipes, greedy decoding. The test also checks the serving process.ONEBIT_HRX=OFF(what CI runs) it's unchanged, and the smoke test passes.Not yet (docs/hrx.md)
🤖 Generated with Claude Code