UPSTREAM PR #19053: CUDA: Fix loop unrolling for BW in mul_mat_q_stream_k_fixup#1107
Open
UPSTREAM PR #19053: CUDA: Fix loop unrolling for BW in mul_mat_q_stream_k_fixup#1107
Conversation
By providing stride_* variables as size_t (i.e., 64-bit) the compiler can correctly unroll the [two for-loops](https://github.com/ggml-org/llama.cpp/blob/557515be1e93ed8939dd8a7c7d08765fdbe8be31/ggml/src/ggml-cuda/mmq.cuh#L3789-L3816) on BW. This gives some perf for prefill/pp phase on BW, while not affecting other SMs: | GPU | Model | Test | t/s master | t/s osimons/fix_bw_mmq_fixup_kernel | Speedup | |:--------------------------------------------------------|:----------------------|:-------|-------------:|--------------------------------------:|----------:| | NVIDIA RTX 6000 Ada Generation | gpt-oss 20B MXFP4 MoE | pp8096 | 8404.05 | 8375.79 | 1.00 | | NVIDIA RTX 6000 Ada Generation | llama 3B Q4_K_M | pp8096 | 16148.93 | 16019.60 | 0.99 | | NVIDIA RTX 6000 Ada Generation | llama 8B Q4_0 | pp8096 | 8008.29 | 7978.80 | 1.00 | | NVIDIA RTX 6000 Ada Generation | nemotron_h 9B BF16 | pp8096 | 4263.16 | 4248.53 | 1.00 | | NVIDIA RTX 6000 Ada Generation | nemotron_h 9B Q4_K_M | pp8096 | 5165.11 | 5157.43 | 1.00 | | NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition | gpt-oss 20B MXFP4 MoE | pp8096 | 12582.80 | 12758.37 | 1.01 | | NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition | llama 3B Q4_K_M | pp8096 | 16879.10 | 17619.47 | 1.04 | | NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition | llama 8B Q4_0 | pp8096 | 10649.90 | 10982.65 | 1.03 | | NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition | nemotron_h 9B BF16 | pp8096 | 7717.73 | 7716.22 | 1.00 | | NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition | nemotron_h 9B Q4_K_M | pp8096 | 7301.90 | 7370.38 | 1.01 |
loci-dev
temporarily deployed
to
PROD__AL_DEMO
GitHub Actions
Inactive
|
No meaningful performance changes were detected across 112622 analyzed functions in the following binaries: build.bin.llama-tts, build.bin.libllama.so, build.bin.llama-cvector-generator, build.bin.libmtmd.so, build.bin.llama-tokenize, build.bin.llama-bench, build.bin.libggml.so, build.bin.libggml-cpu.so, build.bin.libggml-base.so, build.bin.llama-gemma3-cli, build.bin.llama-gguf-split, build.bin.llama-llava-cli, build.bin.llama-minicpmv-cli, build.bin.llama-quantize, build.bin.llama-qwen2vl-cli. 🔎 Full breakdown: Loci Inspector. |
loci-dev
force-pushed
the
main
branch
26 times, most recently
from
d613f70 to
6a853c2
Compare
loci-dev
force-pushed
the
main
branch
30 times, most recently
from
048ad94 to
6c1fde6
Compare
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
2 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.
Note
Source pull request: ggml-org/llama.cpp#19053
By providing stride_* variables as
size_t(i.e., 64-bit), the compiler can correctly unroll the two for-loops on BW. This gives some perf for prefill/pp phase on BW, while not affecting other SMs.For pointer arithmetic inside loops, general performance guidance moving forward is likely to be to perform it in 64-bit unless strictly necessary.
Perf numbers