Repository navigation
Vulkan from upstream llama.cpp: its own pin, kept on the latest release - #28
Merged
Merged
Conversation
third_party/llama.cpp stays on the llama.cpp + hrx-system pair AMD tests for HRX. That pair lags upstream (f1a0aca is 928 commits behind and cannot load Qwen3.8-Flash-Next, qwen4exp), so the Vulkan route now has its own tree: - third_party/llama.cpp-vulkan: ggml-org/llama.cpp v0.5.0 (7fe450e19305). - ONEBIT_VULKAN (cmake/vulkan.cmake) builds its llama-server; 1bit serve --device vulkan prefers it, --device hrx keeps the HRX build, and without ONEBIT_VULKAN vulkan falls back to the HRX build as before. - bump-llama-vulkan.yml opens a PR daily when upstream publishes a release. - docs/vulkan.md; serve.md, hrx.md, README and NOTICE updated. Verified on Strix Halo: ctest 3/3 (serve_e2e_vulkan_upstream PASS); Qwen3.8-27B UD-Q4_K_XL and Qwen3.8-Flash-Next UD-Q4_K_XL both answer "The capital of France is Paris." through 1bit serve (12.2 / 22.0 tok/s); the HRX pin fails to load Flash-Next. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
bong-water-water-bong
pushed a commit
that referenced
this pull request
Sep 27, 2026
…bal barrier (#123, #140) Brings 1bit-MONSTER/llama.cpp#28 and #29: in every decode-split reduce_fused variant the barrier between the reduce's global output stores and pack_completed_q8's output loads was kernel.barrier<workgroup> (LDS only), so the next_q8 copy could be packed from stale output. It is now kernel.barrier<global> followed by the LDS barrier (5 sites, ops and qwen3_moe corpora; #29 restores the LDS fence #28 dropped and adds the missing global barrier to the undispatched standalone reduce_f32 kernels). The pin moves fa226f9 -> 00adc2b, fast-forward. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
bong-water-water-bong
added a commit
that referenced
this pull request
Sep 27, 2026
…bal barrier (#123, #140) (#176) Brings 1bit-MONSTER/llama.cpp#28 and #29: in every decode-split reduce_fused variant the barrier between the reduce's global output stores and pack_completed_q8's output loads was kernel.barrier<workgroup> (LDS only), so the next_q8 copy could be packed from stale output. It is now kernel.barrier<global> followed by the LDS barrier (5 sites, ops and qwen3_moe corpora; #29 restores the LDS fence #28 dropped and adds the missing global barrier to the undispatched standalone reduce_f32 kernels). The pin moves fa226f9 -> 00adc2b, fast-forward. 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.
third_party/llama.cppstays on the llama.cpp + hrx-system pair AMD tests for HRX. That pair lags upstream:f1a0acais 928 commits behind and cannot load Qwen3.8-Flash-Next (qwen4exp, added upstream 2026-08-27). The Vulkan route now gets its own pin, so new architectures land there the day upstream releases them.third_party/llama.cpp-vulkan: ggml-org/llama.cpp v0.5.0 (7fe450e19305).ONEBIT_VULKAN(cmake/vulkan.cmake) builds itsllama-server.1bit serve --device vulkanprefers it;--device hrxkeeps the HRX build; withoutONEBIT_VULKAN, vulkan falls back to the HRX build as before.bump-llama-vulkan.yml: a daily PR when upstream publishes a release (releases, not master: each one passed upstream CI).docs/vulkan.md; serve.md, hrx.md, README and NOTICE updated.--device vulkanbump-llama-vulkan.yml--device hrxbump-hrx.ymlVerified on Strix Halo:
ONEBIT_VULKAN=ON, Qwen3-0.6B)serve_e2e_vulkan_upstreamPASS1bit serve --device vulkan--ctx-size 8192🤖 Generated with Claude Code