CUDA: enable FA for FP32 KV cache #16546
Merged
Conversation
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
github-actions
bot
added
Nvidia GPU
ggerganov
approved these changes
Oct 14, 2025
|
Slightly off-topic, but would it be doable to add BF16 support for applicable hardware? |
|
If we're just talking about the bare minimum support as was done in this PR the answer is yes, if you mean with actually good performance the answer is that it depends on the hardware. For NVIDIA you get BF16 tensor cores with Ampere or newer but for hardware-accelerated basic BF16 operations like addition or multiplication you need Hopper or newer (though I would expect internally those are mapped to FP32 instructions). In my opinion it would not be worth the effort to implement support in the kernels themselves. |
ddh0
added a commit
to ddh0/llama.cpp
that referenced
this pull request
Oct 14, 2025
* cuda : remove legacy copy-op pointer indirection code (ggml-org#16485) * remove legacy copy-op pointer indirection code * further removal of copy-op indirection code * renamed check_node_graph_compatibility_and_refresh_copy_ops function * CUDA: add fp kernel for larger batch size MoE (ggml-org#16512) * CUDA: kernel for larger batch sizes for MoE * WIP * WIP * WIP * WIP * WIP * WIP * fixup * tests * Move mmq_ids_helper to mmid * cleanup * Remove redundant checks * CUDA: use fastdiv + ggml_cuda_mad for mmvf (ggml-org#16557) * CUDA: use fastdiv + ggml_cuda_mad for mmvf * use bf16 directly + fix formatting * Add exception for HIP code * CUDA: enable FA for FP32 KV cache (ggml-org#16546) * vulkan: Improve build time for MSVC (ggml-org#16545) Enable CMP0147 so custom build steps (invoking vulkan-shader-gen) are run in parallel. Enable /MP so source files are compiled in parallel. * vulkan: Support FA with K/V in F32 (ggml-org#16543) * CUDA + openCL: fix bug in accessing rms_norm->src while doing fusion (ggml-org#16577) * vulkan: Add ACC_TYPE_VEC2 implementation (ggml-org#16203) Signed-off-by: Stefan Savic <[email protected]> Co-authored-by: Stefan Savic <[email protected]> * metal : avoid using Metal's gpuAddress property (ggml-org#16576) * metal : avoid using Metal's gpuAddress property * metal : fix rope kernels buffer check --------- Signed-off-by: Stefan Savic <[email protected]> Co-authored-by: Anav Prasad <[email protected]> Co-authored-by: Aman Gupta <[email protected]> Co-authored-by: Johannes Gäßler <[email protected]> Co-authored-by: Jeff Bolz <[email protected]> Co-authored-by: SavicStefan <[email protected]> Co-authored-by: Stefan Savic <[email protected]> Co-authored-by: Georgi Gerganov <[email protected]>
yael-works
pushed a commit
to yael-works/llama.cpp
that referenced
this pull request
Oct 15, 2025
gabe-l-hart
added a commit
to gabe-l-hart/llama.cpp
that referenced
this pull request
Oct 15, 2025
* origin/master: Add server-driven parameter defaults and syncing (ggml-org#16515) metal: optimise `GGML_OP_SUM` (ggml-org#16559) server : fix img token logs (ggml-org#16595) llama-quant: add support for mmproj (ggml-org#16592) CUDA: Changing the CUDA scheduling strategy to spin (ggml-org#16585) server : fix mtmd checkpoints (ggml-org#16591) metal : avoid using Metal's gpuAddress property (ggml-org#16576) vulkan: Add ACC_TYPE_VEC2 implementation (ggml-org#16203) CUDA + openCL: fix bug in accessing rms_norm->src while doing fusion (ggml-org#16577) vulkan: Support FA with K/V in F32 (ggml-org#16543) vulkan: Improve build time for MSVC (ggml-org#16545) CUDA: enable FA for FP32 KV cache (ggml-org#16546) CUDA: use fastdiv + ggml_cuda_mad for mmvf (ggml-org#16557) CUDA: add fp kernel for larger batch size MoE (ggml-org#16512) cuda : remove legacy copy-op pointer indirection code (ggml-org#16485) server : dynamic token limit for prompt cache (ggml-org#16560)
Nexesenex
added a commit
to Nexesenex/croco.cpp
that referenced
this pull request
Oct 16, 2025
This reverts commit 9c7185d.
Nexesenex
added a commit
to Nexesenex/croco.cpp
that referenced
this pull request
Oct 16, 2025
This reverts commit 9c7185d.
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.
For #16531 .
Adds CUDA FlashAttention support for FP32 KV cache. The FP32 data is converted to FP16 for the kernel which is inefficient but simple.