vulkan: add Flash Attention support for BFloat16 KV cache.#23420
Merged
Conversation
0cc4m
changed the title
jeffbolznv
reviewed
May 20, 2026
github-actions
Bot
added
Vulkan
jeffbolznv
reviewed
May 21, 2026
jeffbolznv
approved these changes
May 26, 2026
0cc4m
force-pushed
the
0cc4m/vulkan-fa-bf16
branch
from
a8595b4 to
48f0c0a
Compare
CISC
approved these changes
May 28, 2026
fewtarius
pushed a commit
to fewtarius/llama.cpp
that referenced
this pull request
May 30, 2026
…3420) * vulkan: add flash attention bf16 kv support * vulkan: bf16 FA coopmat1 support * vulkan: bf16 FA coopmat2 support * fix FA bf16 f32 fallback * fix FA bf16 coopmat1 shader * fix FA bf16 coopmat2 shader * code cleanup * cleanup comment change * address feedback * add O_TYPE for cm2 FA * use O_TYPE for gqaStore function * reduce BFLOAT16 ifdefs
o7si
added a commit
to o7si/llama.cpp
that referenced
this pull request
May 31, 2026
…wercase * upstream/master: (27 commits) vocab : add tokenizer support for jina-embeddings-v2-base-zh (ggml-org#18756) ui: fix ETag truncation with MSVC compiler (ggml-org#23917) docs : update ZenDNN docs for Q8 support (ggml-org#23791) llama: only use one iGPU device by default (ggml-org#23897) webui: add custom CSS injection via config (ggml-org#23904) Support `-fa auto` in llama-bench (ggml-org#23714) opencl: support bf16 by converting to f16 (ggml-org#23839) ui: exclude generated build dirs from prettier and eslint so lint errors stop being masked (ggml-org#23910) TP: fix granularity for Qwen 3.5/3.6 + 3 GPUs (ggml-org#23843) metal : restore im2col implementation for large kernels (ggml-org#23901) test: (test-llama-archs) log the config name first (ggml-org#23885) ci : update ios-xcode release job to macos-26 (ggml-org#23906) ggml : add some lsx support (ggml-org#23798) vulkan: add Flash Attention support for BFloat16 KV cache (ggml-org#23420) ci : fix s390x release job (ggml-org#23898) ci : clear cache instead of "no timestamp" keys + fix macos (ggml-org#23895) llama : do not skip iGPU when only RPC devices are present (ggml-org#23868) server: in SSE mode, send HTTP headers when slot starts (ggml-org#23884) ggml-webgpu: Check earlier for WebGPU required features (ggml-org#23879) ggml-webgpu: add q4_0/q8_0 SET_ROWS (ggml-org#23760) ... # Conflicts: # gguf-py/gguf/vocab.py # src/llama-vocab.cpp
turbo-tan
pushed a commit
to turbo-tan/llama.cpp-tq3
that referenced
this pull request
Jun 2, 2026
…3420) * vulkan: add flash attention bf16 kv support * vulkan: bf16 FA coopmat1 support * vulkan: bf16 FA coopmat2 support * fix FA bf16 f32 fallback * fix FA bf16 coopmat1 shader * fix FA bf16 coopmat2 shader * code cleanup * cleanup comment change * address feedback * add O_TYPE for cm2 FA * use O_TYPE for gqaStore function * reduce BFLOAT16 ifdefs
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
3 participants
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.
Overview
This PR adds FA support for symmetrical use of bfloat16 kv cache in the Vulkan backend, meaning it only supports both k and v in bfloat16 format. Because there is no general arithmetic support for bfloat16, the non-coopmat path uses the scalar float32-fallback path.
Requirements