* Remove redundant CUDA copies after gated_delta_net.
Currently, GDN writes recurrent state snapshots into its output tail, then the graph immediately copies those snapshots into ssm_states_all. With MTP draft length 3, target decode uses K=4, so that becomes 4 extra ggml_cuda_cpy calls.
The change detects that gated_delta_net -> view -> cpy pattern and makes the CUDA GDN kernel write the state snapshot(s) directly into the recurrent cache, skipping the intermediate tail writes and copy kernels when safe.
* Address review comments
* [CUDA] Added a cudaMemcpy2DAsync fast path to ggml_cuda_cpy
Add a CUDA ggml_cpy fast path for same-type, same-shape strided copies that are just 2D pitched block copies.
When tensors are not fully contiguous but each row is contiguous, it now uses cudaMemcpy2DAsync instead of the slow element-wise scalar copy kernel.
This fixes the GDN recurrent snapshot update with -np 4, where rollback slots are separated by cache stride gaps.
* Add new tests that execute the new optimized strided copy path
* Return unsupported for strided copy in OpenVINO, as new tests are failing
* Make ggml_gated_delta_net take only the initial recurrent state (D, 1, n_seqs) and passes the snapshot count K as an op parameter instead of inferring it from state->ne[1].
Remove the padding hack and copy all emitted snapshots into the recurrent cache with a single strided ggml_cpy
* Make GDN changes in all backends. Address review comments.
* Fix CI build errors
* Support `-fa auto` in llama-bench
Make the default value of `-ngl` -1, similar to other tools.
Update README with latest usage and examples
* Address review comments
* Move to backend sampling for MTP draft path
Run top_k(10) on the draft backend. D2H transfers happen only for the top 10 logits
Make backend sampling more robust and fallback to CPU on failure cases, such as with "-sm tensor" or when a backend doesn't support TOP_K.
* Allow sampler chains to be partially offloaded to backend
* Add --spec-draft-backend-sampling argument. Enabled by default.
* Write a readme on Multi-GPU usage in llama.cpp
* Apply suggestions from code review
Co-authored-by: Johannes Gäßler <johannesg@5d6.de>
* Address review comments
* Apply suggestions from code review
Co-authored-by: Johannes Gäßler <johannesg@5d6.de>
---------
Co-authored-by: Johannes Gäßler <johannesg@5d6.de>
The previous code worked only for full tensor reads and writes and was hitting `GGML_ASSERT(size == ggml_nbytes(tensor)); ` assert when tested with llama-server.
* Fix delayed AllReduce on Gemma-4 MoE
Skip forward past nodes that don't consume the current one, and allow a chain of MULs.
* Check for all sources before skipping nodes
* Address review comments
* cache subgraph splits when cgraph is unchanged
Skip per-call subgraph construction in ggml_backend_meta_graph_compute when the same ggml_cgraph is used consecutively.
Assign uid to every sub-graph so that CUDA's fast uid check path hits too.
* Address review comments
* Keep the scope as is
* Rename last_uid and last_n_subgraphs field. Remove last_max_tmp_size field. Refactor code.
* Address review comments
* Update ggml/src/ggml-backend-meta.cpp
Co-authored-by: Johannes Gäßler <johannesg@5d6.de>
* Update ggml/src/ggml-backend-meta.cpp
Co-authored-by: Johannes Gäßler <johannesg@5d6.de>
---------
Co-authored-by: Johannes Gäßler <johannesg@5d6.de>
* Write an optimized flash_attn_stream_k_fixup kernel
Write a specialized and more optimized kernel for cases where nblocks_stream_k is multiple of ntiles_dst.
Make nblocks_stream_k to multiple of ntiles_dst if nblocks_stream_k > 2 * ntiles_dst
* Use the new kernel only for nblocks_stream_k_raw > 4 * ntiles_dst to make sure we have enough concurrency on GPUs
* Address review comments
* Address review comments
* Revert variable names to original
* Optimize MOE GEMV kernel for BS > 1.
The previous MOE kernel for BS > 1 had too many thread blocks (nrows_x, nchannels_dst, ncols_dst), with very little work per block. block of (32, 4) was doing inner dot product for a single row.
New mul_mat_vec_q_moe kernel is dedicated for MoE multi-token kernel with grid (ceil(nrows_x/rpb), nchannels_dst), block (warp_size, ncols_dst). Each warp handles two rows independently with warp-level reduction only (no shared memory sync).
This change doesn't increase any compilation time as a single template instance is needed per type. This also simplifies the original GEMV kernel and gets rid of `is_multi_token_id` specialization.
* Remove em-dashes
* Cherry-pick changes from @am17an PR https://github.com/ggml-org/llama.cpp/pull/20885 to enable small_k optimization only for cases where it benefits
Increase max batch size for MMVQ kernels for MUL_MAT_ID to 8
* Make the max batch size for MOE GEMV kernel configurable based on GPU arch and datatype
---------
Co-authored-by: Aman Gupta <amangupta052@gmail.com>
* Increase per-thread work if the K-dimension is small
With tensor parallelism, the K-dimension of the FFN-down matrices is split, which makes it quite small, especially for MOEs. For example, Qwen3-30b-A3B has a K-dimension of 768, and Qwen3235B-A22B has k-dimension of 1536.
The current heuristic uses a group of 4 warps irrespective of K-dimension size, resulting in some of the threads being idle. This results in poor performance for these matrices.
This change increases the number of output elements per block for such cases.
* Limit this change to ncols_dst = 1
* tab to space
* Improve CUDA graph capture
Currently, CUDA graphs are eagerly enabled on the first call to ggml_backend_cuda_graph_compute. If the graph properties keep changing (4+ consecutive updates), the graph is permanently disabled. This is suboptimal because:
- The first call always incurs CUDA graph capture overhead even if the graph is unstable
- Once permanently disabled, CUDA graphs never re-enable even after the graph stabilizes (e.g., switching from prompt processing to decode)
The new approach delays CUDA graph activation until warmup completes: the same cgraph must be called at least twice with matching properties before CUDA graph capture begins. This avoids wasted capture overhead on volatile graphs and allows graphs to become eligible once they stabilize.
This also fixes issues such as https://github.com/ggml-org/llama.cpp/discussions/19708
* Update ggml/src/ggml-cuda/ggml-cuda.cu
Co-authored-by: Johannes Gäßler <johannesg@5d6.de>
* Remove EM dashes
* Update ggml/src/ggml-cuda/ggml-cuda.cu
Co-authored-by: Aman Gupta <amangupta052@gmail.com>
---------
Co-authored-by: Johannes Gäßler <johannesg@5d6.de>
Co-authored-by: Aman Gupta <amangupta052@gmail.com>
Hangs were reported on Jetson Orin AGX if we set CUDA_SCALE_LAUNCH_QUEUES=4x. Reverting the previous PR (#19042) and updating the document to consider setting CUDA_SCALE_LAUNCH_QUEUES=4x for faster throughput on multi-GPU systems.
* [CUDA] Reduce CPU-side stalls due to the CUDA command buffer being full
With pipeline parallelism, during prompt processing, the CPU-side CUDA command buffer gets full, stalling the CPU. Due to this, enough work doesn't get submitted to the GPU, causing bubbles in the GPU timeline.
Fix this by setting the CUDA environment variable CUDA_SCALE_LAUNCH_QUEUES to 4x to increase the command buffer size.
* Set the env variable in the CUDA backend registry allocation
* Add link to PR in code comment
* Remove warning logs and update documentation
