CUDA: fuse RMS_NORM + SCALE into one kernel - #29393
Merged
Merged
Conversation
- ggml-org#28068 builds the GDN q/k l2norm as ggml_scale(ggml_rms_norm(x, eps/n), 1/sqrt(n)). This adds 2 SCALE nodes per GDN layer, 96 extra kernel launches per ubatch on Qwen3.8-27B (48 GDN layers). - The extra kernels take no measurable GPU time, but each launch has a host/driver cost. It is small with plain batch processing and about 10x larger with draft-mtp speculative decoding. - rms_norm_f32 gets a do_scale flag, the same pattern as do_multiply/do_add, so the fused path shares the kernel, the reduction and the launcher. It computes scale * (rsqrt(mean + eps) * x), which matches the unfused rms_norm + scale bit for bit, so ggml-org#28068 numerics are kept. - Fusion only fires when SCALE has no bias and the rms_norm output has a single consumer (ggml_can_fuse). - Metal (ggml-org#28948) and SYCL (ggml-org#28931) already fuse the same pattern. Measured on 2x GTX 1080 Ti (sm_61, PCIe 3.0 x16 + x4), i7-13700KF, Windows 11, driver 582.66, CUDA 12.9. Qwen3.8-27B-UD-Q4_K_XL, -ngl 99 -ts 53,47 -ot token_embd=CPU, master fee39dd. llama-bench -ub 128,512 -p 512,2048 -n 128 -r 5, tok/s: build pp512@128 pp2048@128 pp2048@512 tg128 master 367.7 419.1 385.4 12.90 master + fix 372.6 420.6 388.6 12.98 +1.3% +0.4% +0.8% +0.6% llama-server cold prefill, -c 56000 -ub 128 -b 2048, draft-mtp n-max 3 p-min 0.5, mean of 2 rounds x 3 reps: build pp 8000 pp 20000 master 356.5 322.0 master + fix 371.4 (+4.2%) 337.4 (+4.8%) - Launches per ubatch go from 1032.9 + 841.7 back to 978.9 + 799.7 (CUDA0 + CUDA1), the b10828 count. The GPU op sum is unchanged. - test-backend-ops RMS_NORM_SCALE, NORM_SCALE, RMS_NORM_MUL_ADD, RMS_NORM_MUL_ROPE, RMS_NORM, RMS_NORM_BACK, NORM, L2_NORM and SCALE all pass on both GPUs. - Perplexity is identical to the unfused build: 3.2030 +/- 0.0559 at -c 2048, 16 chunks. - Draft acceptance counts per request match the unfused build. Assisted-by: Claude Opus 5.5
|
Hi @InflexCZE, thanks for your contribution! Per our contribution guidelines, the automated PR checker found the following issue(s) that need your attention:
Please note that maintainers reserve the right to make final decisions on PRs. If you believe there is a mistake, please comment below. |
ggml-gh-bot
Bot
added
the
draft
am17an
approved these changes
Sep 24, 2026
am17an
added
the
merge ready
github-actions
Bot
added
ggml
Contributor
Author
|
@ggml-gh-bot Fixed contributing guidelines link. Was that it? |
am17an
marked this pull request as ready for review
September 25, 2026 02:47
adromir
pushed a commit
to adromir/llama-cpp-turboquant
that referenced
this pull request
Sep 25, 2026
- ggml-org#28068 builds the GDN q/k l2norm as ggml_scale(ggml_rms_norm(x, eps/n), 1/sqrt(n)). This adds 2 SCALE nodes per GDN layer, 96 extra kernel launches per ubatch on Qwen3.8-27B (48 GDN layers). - The extra kernels take no measurable GPU time, but each launch has a host/driver cost. It is small with plain batch processing and about 10x larger with draft-mtp speculative decoding. - rms_norm_f32 gets a do_scale flag, the same pattern as do_multiply/do_add, so the fused path shares the kernel, the reduction and the launcher. It computes scale * (rsqrt(mean + eps) * x), which matches the unfused rms_norm + scale bit for bit, so ggml-org#28068 numerics are kept. - Fusion only fires when SCALE has no bias and the rms_norm output has a single consumer (ggml_can_fuse). - Metal (ggml-org#28948) and SYCL (ggml-org#28931) already fuse the same pattern. Measured on 2x GTX 1080 Ti (sm_61, PCIe 3.0 x16 + x4), i7-13700KF, Windows 11, driver 582.66, CUDA 12.9. Qwen3.8-27B-UD-Q4_K_XL, -ngl 99 -ts 53,47 -ot token_embd=CPU, master fee39dd. llama-bench -ub 128,512 -p 512,2048 -n 128 -r 5, tok/s: build pp512@128 pp2048@128 pp2048@512 tg128 master 367.7 419.1 385.4 12.90 master + fix 372.6 420.6 388.6 12.98 +1.3% +0.4% +0.8% +0.6% llama-server cold prefill, -c 56000 -ub 128 -b 2048, draft-mtp n-max 3 p-min 0.5, mean of 2 rounds x 3 reps: build pp 8000 pp 20000 master 356.5 322.0 master + fix 371.4 (+4.2%) 337.4 (+4.8%) - Launches per ubatch go from 1032.9 + 841.7 back to 978.9 + 799.7 (CUDA0 + CUDA1), the b10828 count. The GPU op sum is unchanged. - test-backend-ops RMS_NORM_SCALE, NORM_SCALE, RMS_NORM_MUL_ADD, RMS_NORM_MUL_ROPE, RMS_NORM, RMS_NORM_BACK, NORM, L2_NORM and SCALE all pass on both GPUs. - Perplexity is identical to the unfused build: 3.2030 +/- 0.0559 at -c 2048, 16 chunks. - Draft acceptance counts per request match the unfused build. Assisted-by: Claude Opus 5.5
feal87
added a commit
to feal87/myllama.cpp
that referenced
this pull request
Sep 25, 2026
Merge upstream commits: - llama: add llama_prec_policy + model-driven W4A4 path (ggml-org#24364) - llama: fix tensor split for fused qkv with uneven K/V head sizes (ggml-org#29294) - metal: split fa kernels into per-dtype libraries (ggml-org#29329) - metal: FWHT kernels for block widths above 512 (ggml-org#29095) - CUDA: fuse RMS_NORM + SCALE into one kernel (ggml-org#29393) - common: extract shared unicode path/string helpers (ggml-org#29415) - common,rpc: simplify fs_create_directory_with_parents() (ggml-org#29432) - rpc: include nb in the get_alloc_size cache key (ggml-org#29283) - [SYCL] support sparse FA (ggml-org#28796) - musa: fix PH1 operator failures and build issues (ggml-org#29193) - HIP: bump HIP_VERSION required for fp8 (ggml-org#29231) - opencl: add q5_k bin kernel (ggml-org#29401) - hexagon: add q5_k quant type support (ggml-org#29123) - hexagon: use DMA for contiguous dim1 CONCAT (ggml-org#29404) - mtmd: fix mel preprocessor in LFM2 audio (ggml-org#29403) - vulkan: fix legacy GLSLC without cooperativeMatrix (ggml-org#29409) - gguf-py: ByteLevel processing defaults bos/eos to False (ggml-org#29422) - gguf-py: TemplateProcessing has final word on add_special_token (ggml-org#29417) Assisted-by: Pi
sky-mighty
pushed a commit
to sky-mighty/llama.cpp
that referenced
this pull request
Sep 26, 2026
- ggml-org#28068 builds the GDN q/k l2norm as ggml_scale(ggml_rms_norm(x, eps/n), 1/sqrt(n)). This adds 2 SCALE nodes per GDN layer, 96 extra kernel launches per ubatch on Qwen3.8-27B (48 GDN layers). - The extra kernels take no measurable GPU time, but each launch has a host/driver cost. It is small with plain batch processing and about 10x larger with draft-mtp speculative decoding. - rms_norm_f32 gets a do_scale flag, the same pattern as do_multiply/do_add, so the fused path shares the kernel, the reduction and the launcher. It computes scale * (rsqrt(mean + eps) * x), which matches the unfused rms_norm + scale bit for bit, so ggml-org#28068 numerics are kept. - Fusion only fires when SCALE has no bias and the rms_norm output has a single consumer (ggml_can_fuse). - Metal (ggml-org#28948) and SYCL (ggml-org#28931) already fuse the same pattern. Measured on 2x GTX 1080 Ti (sm_61, PCIe 3.0 x16 + x4), i7-13700KF, Windows 11, driver 582.66, CUDA 12.9. Qwen3.8-27B-UD-Q4_K_XL, -ngl 99 -ts 53,47 -ot token_embd=CPU, master fee39dd. llama-bench -ub 128,512 -p 512,2048 -n 128 -r 5, tok/s: build pp512@128 pp2048@128 pp2048@512 tg128 master 367.7 419.1 385.4 12.90 master + fix 372.6 420.6 388.6 12.98 +1.3% +0.4% +0.8% +0.6% llama-server cold prefill, -c 56000 -ub 128 -b 2048, draft-mtp n-max 3 p-min 0.5, mean of 2 rounds x 3 reps: build pp 8000 pp 20000 master 356.5 322.0 master + fix 371.4 (+4.2%) 337.4 (+4.8%) - Launches per ubatch go from 1032.9 + 841.7 back to 978.9 + 799.7 (CUDA0 + CUDA1), the b10828 count. The GPU op sum is unchanged. - test-backend-ops RMS_NORM_SCALE, NORM_SCALE, RMS_NORM_MUL_ADD, RMS_NORM_MUL_ROPE, RMS_NORM, RMS_NORM_BACK, NORM, L2_NORM and SCALE all pass on both GPUs. - Perplexity is identical to the unfused build: 3.2030 +/- 0.0559 at -c 2048, 16 chunks. - Draft acceptance counts per request match the unfused build. Assisted-by: Claude Opus 5.5
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
Adds Cuda backend fusion optimizations for MoE / Gated Delta Net paths
Fixes ~4% regression in prefil from models: fix GDN normalization from
maxtorsqrt#28068The same fusion, already merged on other backends:
Affected architectures: qwen35, qwen35moe, qwen3next, qwen4exp, kimi-linear, kimi-k3, bailingmoe3.
Additional information
llama-server -m Qwen3.8-27B-UD-Q4_K_XL.gguf -ngl 99 -np 1 -c 56000 -ub 128 -b 2048 -ts 53,47 -ot "token_embd\.weight=CPU" --spec-type draft-mtp --spec-draft-n-max 3 --spec-draft-p-min 0.5 --cache-ram 20480llama-bench -m Qwen3.8-27B-UD-Q4_K_XL.gguf -ngl 99 -ts 53/47 -ot "token_embd\.weight=CPU" -b 2048 -ub 128,512 -p 512,2048 -n 128 -r 5 (No speculative decoding)Correctness
llama-perplexity -c 2048 --chunks 16 -ub 128ggml_l2_norm)test-backend-ops test -b CUDA0maxtorsqrt#28068 (b10829) adds 96 launches per ubatch (2 per GDN layer x 48 layers).Scope and limits
Requirements