Commit Graph
17 Commits
Author SHA1 Message Date
Oliver Simons 1a064ab092 CUDA: Improve NVFP4 W4A4 activation quantization (#25730)
* Squash history before conflict-resolution during rebase on master

WIP commit

Add 32-byte loads, restore per-block amax

Use nvfp4x4 intrinsic when available

Fuse per-channel amax and quantization kernels

Do pointer arithmetic only once on x

Remove unnecessary ternary in the load

We assert on host side that ne00 is 64-aligned

Add back scale-search, but optimize it with intrinsics

Code cleanup

Make scale in MMQ-epilogue NVFP4-specific/restrictive for now

Remove unneeded include, add comment

Fix trailing whitespace

Guard __builtin_align__(32) struct to NVIDIA

Seems like HIP doesn't have this available, see https://github.com/ggml-org/llama.cpp/actions/runs/29438651734/job/87431623001

* compiler massaging to avoid unnecessary LDCs

* kvalues_mxfp4 -> kvalues_nvfp4 in quantize_mmq_nvfp4

* Always pass in src1_scale.ptr

* Extract ggml_cuda_is_aligned helper
2026-07-22 19:28:02 +02:00
Pranesh Gonegandlaandpraneshgo 5839ba3524 CUDA: dedup MoE gate/up activation quantization (#25441)
* CUDA: dedup MoE gate/up activation quantization (fp4)

For MoE gate/up projections the src1 activation is broadcast across the
routed experts (ne11 == 1), so ids_src1 maps every one of a token's
n_expert_used slots to the same physical row. The MMQ path therefore
re-quantized each token's activation n_expert_used times.

For fp4 (NVFP4/MXFP4) src0, quantize each unique token row once instead of
once per expert. For NVFP4 a single quantize+scatter kernel
(quantize_scatter_mmq_nvfp4) quantizes each token once and writes the
resulting block_fp4_mmq straight to all n_expert_used slots, using an
inverse token->compact-row map (build_tok2c). MXFP4, and
GGML_CUDA_MOE_QUANT_GATHER=1, use a two-kernel variant: quantize unique
rows then gather into the expert-sorted layout (gather_mmq_fp4_blocks).
Both are bit-identical to the previous gather-then-quantize path (identical
source data, deterministic per-block quantization), verified by
test-backend-ops MUL_MAT_ID (type_a=nvfp4, broadcast b=1; 790/790 for the
default, gather, and per-expert paths) and by coherent end-to-end
generation. Set GGML_CUDA_NO_MOE_QUANT_DEDUP=1 to force the original
per-expert path.

Same-binary A/B on RTX 5090 (sm_120), Qwen3.6-35B-A3B-NVFP4 prefill @8192
(nsys, graphs-off; the unchanged mul_mat_q GEMM confirms stable clocks):
activation-quant GPU-busy drops 61% (78.2 -> 30.4 ms) with the fused
quantize+scatter, vs 33% (78.2 -> 52.8 ms) for the two-kernel gather. The
fused path avoids materializing and re-reading the 8x compact buffer,
writing the expert copies directly from registers.

* CUDA: bounds-check token ids in build_tok2c_kernel

Guard against malformed ids_src1: skip out-of-range token ids (t < 0 or
t >= n_tokens) and drop entries beyond n_expert_used per token instead of
writing past the token's tok2c region. No behavior change for valid MoE
routing data; test-backend-ops MUL_MAT_ID 790/790.

* Refactor the code based on review comments

- Removed previously added kernels that were not necessary anymore\
- Added an inverse mapping from (token, slot) to compact row. Each token is quantized once and scattered to its compact rows.

* Adding q8_1 support for dedup and addressing review comments

* Add pragma unrolls

* Remove redundant cudaMemsetAsync call

* Removing follow up redundancies

---------

Co-authored-by: praneshgo <227579474+praneshgo@users.noreply.github.com>
2026-07-16 09:02:25 +02:00
leonardHONG f6f12e43fa CUDA: tighter MMQ src1 buffer size for native fp4 (#25613) 2026-07-15 23:21:22 +08:00
Andreas KieslingerandOliver Simons 9e58d4d692 Avoid PDL race conditions by disabling __restrict__ when PDL is used (#24030)
* Removes __restrict__ from PDL kernel headers due to incompatibility with
PDL. Adds preprocessor directives based on arch in kernel body to add
__restrict__ to retain performance on older architectures.

* Simplifies new __restrict__ usage via macro

* Add hopper to PDL __restrict__ fix.

Co-authored-by: Oliver Simons <osimons@nvidia.com>

---------

Co-authored-by: Oliver Simons <osimons@nvidia.com>
2026-06-03 13:56:42 +02:00
Andreas Kieslinger e947228222 Programmatic Dependent Launch (PDL) for more performance on newer NVIDIA GPUs (Hopper+) (#22522)
* Adds initial PDL setup.

* Adds PDL barriers based on simple heuristic: place "sync" before first input pointer access, and "launch" after last write, e.g. to tensors like dst.

* Further optimization pass of the first half of kernels

* Optimized PDL barriers for the second batch of kernels

* Further refinements after rebase.

* Moves pdl logic to separate function, removes some whitespace

* Strips post-hoc PDL logic

* Adds stream capture PDL setup. Enrolls quantize_q8_1 to leverage pdl to
overlap execution with previous kernels

* Enrolls mul_mat_vec_q, rms_norm_f32 and k_bin_bcast (partly) into PDL

* Enrolls mmvf, rope, set-rows and topk kernels for gpt-oss into PDL

* Introduce ggml_cuda_kernel_launch, to abstract away cudaLaunchKernelEx,
to enable hip/musa compatibility

* Enrolls cpy_scalar_contiguous, k_get_rows_float and rms_norm_f32

* Enrolls flash_attn_combine_results

* Fix: Drops needless and broken check of CUDA arch for PDL. PDL either
works or is without effect.

* Enrolls flash-attention kernels to pdl

* Fix: inlines ggml_cuda_kernel_launch, and uses perfect forwarding for
kernels args. This fixes PDL.

* Perf: Enrolls k_bin_bcast variadic template invocation into PDL, via
and template alias and template expansion

* Enrolls all remaining kernels for qwen3-coder-next into PDL

* Remove all PDL LC calls to create a baseline

* Added LC according to internal guidance and tested kernel performance.

* Enrols missing qwen3-5 kernels passively into PDL.

* Kernel optimizations (LC signals) for qwen3.5

* Enrolls ssm-scan kernels into PDL

* Adds GGML_CUDA_PDL command line option to toggle PDL.

* Fix: Ada and lower compilation by guarding PDL calls correctly

* Cleanup: Removes commented out GGML_CUDA_PDL_LC

* Cleanup: Removes experimental comments

* Adds 90-virtual to build script so that Hopper GPUs can leverage PDL.

* Adds stricter checks to enable PDL, adds env-check to disable it, and removes now superfluous compile option to enable PDL.

* Fix: Correct PDL en/disablement based on device-side arch check. Host
side check is UB. Required moving from macros to inlined functions

* Fix: default-disable PDL. Enable by setting GGML_CUDA_ENABLE_PDL=1

* Enable PDL by default for Hopper+ devices

* Enrolls softcap_f32 and two flash_attn kernels into PDL.

* Improves flash attn PDL barrier placement

* Fix: Perf regression on ada; excludes ada and below from PDL launches

* Improves some sync barrier placements

* Drops superfluous constructor

* Adds #endif guard comments

* Reverts experimental change to top-k-moe.cu, which moved expensive allocations
in front of the PDL barrier. It did not have a meaningful impact.

* Exchanges GGML_CUDA_DISABLE_PDL with GGML_CUDA_PDL. IFF GGML_CUDA_PDL=0
PDL is disabled

* Revert "Drops superfluous constructor". Adds const to remaining
arguments

This reverts commit 12b1d250da0089ae02a9bb71bbb3fd6d70f6f2f1.

* Cleanup: Removes and fixes some comments and whitespace

* Clarifies comment of sync-barrier position

* Relocates and refactors PDL launch functions and accessories

* Adds error checking to the regular kernel launch path

* Drops "auto" in favor of "ggml_cuda_kernel_params"

* Adds "const" to ggml_cuda_kernel_launch_params

* [Whitespace] Adds final newline to common.cuh to make editorconfig CI job happy
2026-05-20 13:59:02 +02:00
Michael Wand fc2b0053ff ggml-cuda: Repost of 21896: Blackwell native NVFP4 support (#22196) 2026-04-29 06:47:42 +08:00
92f7da00b4 chore : correct typos [no ci] (#20041)
* fix(docs): correct typos found during code review

Non-functional changes only:
- Fixed minor spelling mistakes in comments
- Corrected typos in user-facing strings
- No variables, logic, or functional code was modified.

Signed-off-by: Marcel Petrick <mail@marcelpetrick.it>

* Update docs/backend/CANN.md

Co-authored-by: Aaron Teo <taronaeo@gmail.com>

* Revert "Auxiliary commit to revert individual files from 846d1c301281178efbc6ce6060ad34c1ebe45af8"

This reverts commit 02fcf0c7db661d5ff3eff96b2b2db9fdb7213256.

* Update tests/test-backend-ops.cpp

Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>

* Update tests/test-backend-ops.cpp

Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>

---------

Signed-off-by: Marcel Petrick <mail@marcelpetrick.it>
Co-authored-by: Aaron Teo <taronaeo@gmail.com>
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>
2026-03-05 08:50:21 +01:00
Aman GuptaandAman Gupta <aman> c8a2417d7b CUDA: experimental native mxfp4 support for blackwell (#17906)
* CUDA: experimental native mxfp4 support for blackwell

* optimize load_tiles

* optimize quantize_mxfp4

* cleanup

* first pass review: formatting

* use interleaved layout for mma

* mmq: add assert for size

* use __nv_fp4x4_e2m1

* use iter_k as 512, cleanup

* Use 1200 as blackwell instead of 1000

* address review comments

* mmq: fix stride

* quantize.cu: use reference impl of e8m0 scale

* address review comments

* add 120f-virtual + minor fixes

---------

Co-authored-by: Aman Gupta <aman>
2025-12-24 22:28:26 +08:00
Johannes Gäßler 5143fa895e CUDA: fastdiv, launch bounds for mmvq + q8_1 quant (#15802)
* CUDA: fastdiv, launch bounds for mmvq + q8_1 quant
2025-09-05 16:07:02 +02:00
Johannes Gäßler 4696d56749 CUDA: fix crash on large batch size for quant. MoE (#13537) 2025-05-14 16:41:02 +02:00
Johannes Gäßler 9070365020 CUDA: fix logic for clearing padding with -ngl 0 (#13320) 2025-05-05 22:32:13 +02:00
Johannes Gäßler e1e8e0991f CUDA: batched+noncont MMQ, refactor bs>1 MoE code (#13199) 2025-04-30 23:12:59 +02:00
Johannes Gäßler 658987cfc9 CUDA: noncont MMVQ + batched bs1 MUL_MAT_ID (#13014)
* CUDA: noncont MMVQ + batched bs1 MUL_MAT_ID

* fix logic for RoPE support, CUDA graphs
2025-04-22 21:27:40 +02:00
Diego DevesaandJohannes Gäßler a5e47592b6 cuda : optimize argmax (#10441)
* cuda : optimize argmax

* remove unused parameter

ggml-ci

* fixup : use full warps

ggml-ci

* Apply suggestions from code review

Co-authored-by: Johannes Gäßler <johannesg@5d6.de>

* fix ub

* ggml : check ne00 <= INT32_MAX in argmax and argsort

---------

Co-authored-by: Johannes Gäßler <johannesg@5d6.de>
2024-11-21 18:18:50 +01:00
slaren 2b1f616b20 ggml : reduce hash table reset cost (#8698)
* ggml : reduce hash table reset cost

* fix unreachable code warnings after GGML_ASSERT(false)

* GGML_ASSERT(false) -> GGML_ABORT("fatal error")

* GGML_ABORT use format string
2024-07-27 04:41:55 +02:00
Johannes Gäßler 808aba3916 CUDA: optimize and refactor MMQ (#8416)
* CUDA: optimize and refactor MMQ

* explicit q8_1 memory layouts, add documentation
2024-07-11 16:47:47 +02:00
Georgi Gerganovandslaren f3f65429c4 llama : reorganize source code + improve CMake (#8006)
* scripts : update sync [no ci]

* files : relocate [no ci]

* ci : disable kompute build [no ci]

* cmake : fixes [no ci]

* server : fix mingw build

ggml-ci

* cmake : minor [no ci]

* cmake : link math library [no ci]

* cmake : build normal ggml library (not object library) [no ci]

* cmake : fix kompute build

ggml-ci

* make,cmake : fix LLAMA_CUDA + replace GGML_CDEF_PRIVATE

ggml-ci

* move public backend headers to the public include directory (#8122)

* move public backend headers to the public include directory

* nix test

* spm : fix metal header

---------

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>

* scripts : fix sync paths [no ci]

* scripts : sync ggml-blas.h [no ci]

---------

Co-authored-by: slaren <slarengh@gmail.com>
2024-06-26 18:33:02 +03:00