c7cddefcbde2ccf1829c55ea57bbee92316c294a
70
Commits
| Author | SHA1 | Message | Date | |
|---|---|---|---|---|
 YiChen LvandAman Gupta
|
d789527482 |
spec : Support Step3.5/3.7 flash mtp3 (#24340)
* add mtp_layer_offset + include nextn flags in graph reuse * add llama_set_mtp_layer_offset + llama_model_n_nextn_layer API * offset head select + require all MTP blocks * speculative multi-head process() * speculative multi-head draft() * gather outputs via inp_out_ids * cleanup * fix core * minor cleanup * merged draft_multi_head into draft() * mtp rename nextn * Apply suggestions from code review Co-authored-by: Aman Gupta <amangupta052@gmail.com> * clean-up comments * fix for multi seq * apply suggestions && chain-heads comment * add a reference for chain_heads discussion --------- Co-authored-by: Aman Gupta <amangupta052@gmail.com> |
||
|
Oliver Simons
|
02810c7aa8 |
Fix and restrict NVFP4 edge-cases in llama-graph (#24331)
* Move post-GEMM MUL required for dequant b4 lora and bias add see https://github.com/ggml-org/llama.cpp/pull/23484 : 1. For lora, I would presume we want fully dequantized values before doing the residuals, but this depends on how the LORAs were generated. Literature tells me LORA happens post-mul but pre-bias add https://github.com/ggml-org/llama.cpp/pull/8332 2. For ModelOPT, bias-add should happen on [fully-dequantized values](https://github.com/NVIDIA/Model-Optimizer/blob/b49f9b9e2d747af992d78a3aa7f10efe5a8847e1/modelopt/torch/quantization/backends/nvfp4_gemm.py#L59-L64) * Restrict build_ffn for NVFP4 to supported combinations |
||
|
|
88a39274ec |
spec: add EAGLE3 speculative decoding support (#18039)
* llama : enable layer input extraction * spec: support eagle3 * eagle3: fix params bug * eagle3: support Gemma4 eagle3 from RedHatAI * eagle3: set sync when get features from target Co-authored-by: tnhnyzc <115956684+tnhnyzc@users.noreply.github.com> * eagle3 : fix ubatch handling in embd_layer_inp extraction and encoder Co-authored-by: Doğaç Eldenk <dogacel@gmail.com> * eagle3: adapt to upstream changes * eagle3: fix rebase issues and adapt to upstream changes * eagle3:exclude the eagle3 arch from test-llama-archs * eagle3: fix editorconfig check failures * eagle3: fix multi-seq issue in d2t vocab mapping * cont : minor style / clean-up * spec : remove `common_speculative_setup_draft_model()` * llama : clean-up unused API * eagle3: set d2t vocab mapping in decode graph * cont : assert layer inputs are configured * hparams : use n_embd_inp instead of n_embd_target_features * eagle3: make output.weight optional and inherit from target model when needed * haparams : generic norm-before-residual param * llama-ext : consistent names * cont : fix * hparams : remove target_hidden_size * cparams : rename output_layer_inp -> embeddings_layer_inp * arch : reuse ATTN_NORM_2 instead of adding new hidden norm * llama : clean-up names * cont : add assert + comment * Update conversion/llama.py Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com> --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: tnhnyzc <115956684+tnhnyzc@users.noreply.github.com> Co-authored-by: Doğaç Eldenk <dogacel@gmail.com> Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com> |
||
|
Aman Gupta
|
04eb4c446d | llama : add Gemma4 MTP (#23398) | ||
|
Aman Gupta
|
166fe29492 |
qwen35: use post-norm hidden state for MTP (#24025)
* qwen35: use post-norm hidden state for MTP * rename pre_norm to nextn * fix step35 |
||
|
Hans Florian
|
bfb4308b05 |
model : support granite multilingual embeddings R2 (ibm-granite/granite-embedding-{97,311}m-multilingual-r2) (#22716)
* Add support for the ibm-granite/granite-embedding-{97m,311m}-multilingual-r2 embedding models:
* Added a version of the gpt4o tokenizer that has a fixed regex (better handling of marks), and different token merging setting for the 97m model
* Reused gemma4 tokenizer for the 311m model
* granite-embedding-*-multilingual-r2 : add support SwiGLU FFN for Granite Embedding Multilingual R2
* added new GGUF key <arch>.hidden_activation (LLM_KV_HIDDEN_ACT) + writer
* added a forward declaration of llm_ffn_op_type to llama-hparams.h
* added llm_ffn_op in hparams
* added LLM_FFN_NONE = 0 sentinel to llm_ffn_op_type (value-initialization), modern-bert: explicitly assigns LLM_FFN_GEGLU before reading GGUF (unchanged).
* centralized hidden_act mapping in llama-model.cpp, added llm_ffn_op_type_from_string() helper, mirroring rope_scaling_type/llama_rope_scaling_type_from_string()
* modern-bert reads the GGUF key (when present) and uses the resulting op in its FFN graph
* Added granite-embedding-{97m,311m}-multilingual-r2 to the converter code
* Added the hashes for the granite embedding multilingual R2 models
* Set the hidden_activation in the GGUF if the field is present in config.json (such as for the granite embedding models)
|
||
|
Sigbjørn Skjæret
|
764f1e64a1 | graph : ensure DS32 kq_mask_lid is F32 (#23864) | ||
|
|
1f0aa2a696 |
model : support for DeepseekV32ForCausalLM with generic DeepSeek Sparse Attention (DSA) implementation (#23346)
* llama : support DeepSeek V3.2 model family (with DSA lightning indexer) * convert : handle DeepseekV32ForCausalLM architecture * ggml : support for f16 GGML_OP_FILL * memory : separate hparams argument in llama_kv_cache constructor * memory : add llama_kv_cache_dsa memory (KV cache + lightning indexer cache) * llama : support for LLM_ARCH_DEEPSEEK32 * model : llama_model_deepseek32 implementation * model : merge two scale operations into one in DSA lightning indexer implementation * chore : remove unused code * model : support NVFP4 in DeepSeek V3.2 Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com> * memory : refactoring TODO Co-authored-by: ggerganov <ggerganov@users.noreply.github.com> --------- Co-authored-by: Stanisław Szymczyk <sszymczy@gmail.com> Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com> Co-authored-by: ggerganov <ggerganov@users.noreply.github.com> |
||
|
Aman Gupta
|
031ddb2e08 |
llama: use f16 mask for FA to save VRAM (#23764)
* llama: use f16 mask for FA * review: add llama_cast + formatting * simplify |
||
|
Aman GuptaandGeorgi Gerganov
|
eef59a7642 |
llama: add llm_graph_input_mtp (#23643)
* llama: add llm_graph_input_mtp * rename input_mtp -> input_token_embd * add TODO about mtmd embedding * cont : clean-up --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> |
||
|
Georgi Gerganov
|
d14ce3dab4 |
llama : MTP clean-up (#23269)
* llama : disable equal splits for recurrent memory with partial rollback * spec : re-enable p-min with MTP drafts * spec : re-enable ngram spec in combination with RS rollback * spec : fix ngram-map-* params * spec : fix acceptance logic in combined ngram + draft configs * graph : fix reuse for combined `token` + `embd` batches * spec : log parameters for each speculative implementation - add LOG_INF in each constructor with implementation type and parameters - extract device string logic into common_speculative_get_devices_str() - move 'adding speculative implementation' log from init into constructors Assisted-by: llama.cpp:local pi * spec : extend --spec-default with ngram-map-k4v Assisted-by: llama.cpp:local pi * minor : fix n_embd log * args : update draft.n_max == 3 + regen docs * spec : relax ngram-mod rejection thold to 0.25 @ 5 low * logs : improve * docs : update speculative decoding CLI argument documentation - Add missing draft model CPU scheduling and tensor override parameters - Update --spec-type to include all available types (excluding draft-eagle3 WIP) - Fix default values to match implementation (n_max=3, n_min=0, p_min=0.0) - Remove deprecated options (spec-draft-ctx-size, spec-draft-replace) - Add environment variables for new parameters Assisted-by: llama.cpp:local pi * arg : step-back on adding k4v to the default spec config * cont : fix name |
||
|
Aman GuptaandGeorgi Gerganov
|
255582687b |
llama + spec: MTP Support (#22673)
* spec: support MTP * fix batch size * rename files * cont : simplify (#7) * MTP: clean-up (#9) * MTP: clean-up * review: use llama_context_type instead of llama_graph_type * review: remove llama_model_has_mtp * review: fix convert issues * convert: fix pycheck * review: formatting * use `mtp-` for identifying mtp models * convert: fix mtp conversion * mtp -> draft-mtp * remove unused llama_arch * add need_embd in speculative * llama: allow partial seq_rm for GDN models for speculative decoding Currently speculative checkpoint needs to restart from a checkpoint after some draft tokens are not accepted, this leads to some wastage in running the target again. This PR adds the ability to rollback upto `draft_max` by storing the GDN intermediates. * fix pending state * vulkan: add GDN partial rollback * meta: extend check to axis 1 * metal: add GDN partial rollback Extend the gated delta net kernel to store intermediate states for partial rollback support on the Metal backend. - Add K (snapshot slot count) as a function constant - Read input state from slot 0 of the 3D state tensor - Write intermediate states to different slots during token loop - For K=1, maintain backward-compatible single-slot behavior Ref: https://github.com/ggml-org/llama.cpp/commit/8c05923630110223669f069af2000e9cf10c02bc Assisted-by: llama.cpp:local pi * delta_net_base: use ggml_pad instead of new_tensor * review: add need_rs_seq * review: rename part_bounded to n_rs * review: deslop comments * review: rename, add asserts * server : adjust checkpoint logic (#11) * server : adjust checkpoint logic * cont : rm asserts * server-context: fix early exit * spec : fix compatibility with n-gram and add TODOs (#13) * metal : cleanup * llama : fix faulty bitwise check in recurrent memory * server : disable RS-based MTP in combination with other spec types * spec : add TODOs * cont : fix comment * cont : update comment * common : fix logic for ngram + mtp compat * llama-memory: enable checkpointing with partial rollback * cont: add test-case for loading into a dirty ctx * llama-memory-recurrent: clear rs_idx in clear * download: fix mtp path * llama-arch: fix enorm op * docs: update docs * conversion: fix type annotations --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> |
||
|
PikaPikachu
|
9db77a020c |
model : refactor QKV into common build_qkv and create_tensor_qkv helpers (#21245)
* model : refactor QKV into common build_qkv and create_tensor_qkv helpers * model : extend build_qkv to bert/mpt/dbrx/olmo/lfm2/nemotron-h/granite-hybrid/gemma3n-iswa/t5-dec and fix wqkv_s |
||
|
Sigbjørn Skjæret
|
f772f6e434 |
model : support NVFP4 tensors for Gemma4 (#21971)
* support nvfp4 tensors for Gemma4 * add wo_s to build_attn * add wo_s to build_attn * fix glm4 |
||
|
Georgi Gerganov
|
4eb19514dd |
kv-cache : support attention rotation for heterogeneous iSWA (#21513)
* kv-cache : support attention rotation for heterogeneous iSWA * cont : remove assert |
||
|
Georgi Gerganov
|
744c0c7310 |
llama : rotate activations for better quantization (#21038)
* llama : rotate activations for better quantization * cont : rotate V more + refactor * cont : rotate caches separately + support non-power-of-2 head sizes * cont : simplify * cont : add reference for V rotation * cont : refactor * cont : support context shift * cont : consolidate * cont : dedup + allow different types for the rotation matrix * cont : add env variable to disable rotation * cont : simplify attn rot kv cache logic + rename env * cont : pre-compute the Hadamard matrices |
||
|
Richard Davison
|
1eea6a2968 | graph : add optional scale parameter to build_lora_mm [no ci] (#20427) | ||
|
Richard DavisonandSigbjørn Skjæret
|
5eae9cb1d9 |
ggml : add NVFP4 quantization type support (#19769)
* WIP: add NVFP4 quantization support * tests * improve NVFP4 dot product implementation performance and fix bad super call * typo * Use nvfp4 kvalues * vulkan : fix NVFP4 shader compilation by including kvalues_mxfp4 lookup table * vulcal and perf fixes * wip * Fix metal * fix vulcan * Rename threshold & fix wrong scale * Fix MOE * Shelf backend implementations (CUDA, Metal, Vulkan, arch-specific SIMD) Remove NVFP4 support from GPU backends and architecture-specific optimized dot products. These should be added in separate PRs so backend specialists can review them independently. Reverted files: - ggml-cuda: common.cuh, convert.cu, mmq.cu/cuh, mmvq.cu, vecdotq.cuh, quantize.cu/cuh, mma.cuh, ggml-cuda.cu, fattn-tile.cuh - ggml-metal: ggml-metal.metal, ggml-metal-device.cpp, ggml-metal-impl.h, ggml-metal-ops.cpp - ggml-vulkan: ggml-vulkan.cpp, all vulkan-shaders/* - ggml-cpu arch: arm/quants.c, x86/quants.c, powerpc/quants.c, s390/quants.c Core NVFP4 support (type definition, CPU fallback dot product, quantization, dequantization, conversion) is retained. * Fix arch-fallback.h: add NVFP4 generic fallback for all platforms After shelving backend-specific SIMD implementations, the generic CPU dot product needs to be aliased on ARM, x86, PowerPC, and s390 platforms that previously relied on arch-specific versions. * quantize: add NVFP4 as a quantization type option * Fix ggml_fp32_to_ue4m3: handle subnormal values Previously, values with ue4m3_exp <= 0 were clamped to 0, causing all small scales to underflow. This made NVFP4 quantization via llama-quantize produce garbage (PPL = 5.8M) since typical transformer weights have amax/6.0 in the range 0.001-0.01, which falls in the UE4M3 subnormal range. Now subnormals are properly encoded as man * 2^-9 (exp=0, man=1..7), matching the decode path in ggml_ue4m3_to_fp32. Result: NVFP4 requantization now produces PPL = 15.25 (vs F16 = 14.33), comparable to Q4_1 (PPL = 15.81) at slightly lower BPW (4.70 vs 5.15). * Restore ARM NEON NVFP4 dot product implementation Restores the optimized ggml_vec_dot_nvfp4_q8_0 for ARM NEON using vqtbl1q_s8 lookup and ggml_vdotq_s32 dot products. tg128 performance: 4.37 t/s (generic) -> 13.66 t/s (NEON) = 3.1x speedup * Optimize ARM NEON NVFP4 dot product: LUT + vpaddq + vfmaq - Add ue4m3_scale_lut[128] to ggml-common.h replacing branch-heavy ggml_ue4m3_to_fp32() in the hot loop - Use vpaddq_s32 for pairwise int32 reduction instead of vaddvq_s32 - Accumulate with vfmaq_f32 into float32x4_t vector accumulators tg128: 8.1 -> 31.0 t/s (3.8x speedup, 77% of Q4_1 speed) * ARM NEON NVFP4: rearrange q8 to match nibble layout Alternative approach: rearrange q8 data to match the NVFP4 lo/hi nibble layout instead of rearranging the looked-up NVFP4 values. Eliminates vcombine_s8(vget_low, vget_low) shuffles. Performance is equivalent (~18.5 t/s) - the bottleneck is the 2x block overhead from QK=16 vs QK=32, not the shuffle instructions. * CPU only backend 64 super-block layout * cleanup * Remove unused LUT * int * exclude NVFP4 from unsupported ops in metal build * remove quantization for now * store scales as native UE4M3, preserve original model bits when possible * Update convert_hf_to_gguf.py Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com> * correct comment * format * reduce duplication and cleanup * Address comments * move detection to prepare_tensors * Use math instead of const * Move * fix comment * Shelf quantize tests * Rebase and move check * cleanup * lint * Update gguf-py/gguf/scripts/gguf_convert_endian.py Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com> * Use fallback quant config * Simplify Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com> * organize * Refactor * Update convert_hf_to_gguf.py Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com> * Update convert_hf_to_gguf.py Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com> * Update convert_hf_to_gguf.py Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com> * add quantize_nvfp4 (required for test_quants.py) * add quantize_nvfp4 (required for test_quants.py) * add quantize_nvfp4 (required for test_quants.py) * fix return type --------- Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com> |
||
|
Sigbjørn Skjæret
|
35bee031e1 | graph : remove redundant scale_w parameter (#20235) | ||
|
Aman GuptaandSigbjørn Skjæret
|
b68d75165a |
llama: Add option to merge gate and exp weights (#19139)
* llama: Add option to merge gate and exp weights * Update convert_hf_to_gguf.py Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com> * Update convert_hf_to_gguf.py Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com> * update constants.py * add gate_up for the all MoE models * convert: simplify merge tensor condition * update constants.py * reduce number of models, add create_tensor_gate_up helper --------- Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com> |
||
|
Tarek DakhranandSigbjørn Skjæret
|
8004f3a8d1 |
model : add tokenizer from LFM2.5-Audio-1.5B (#19687)
* model : Add tokenizer from LFM2.5-Audio-1.5B [LFM2.5-Audio-1.5B](https://huggingface.co/LiquidAI/LFM2.5-Audio-1.5B) introduced lightweight audio tokenizer. Tokenizer based on LFM2 architecture and acts as "embedding" model with different input `n_embd` and output `n_embd_out`. To be used in https://github.com/ggml-org/llama.cpp/pull/18641. To convert use ```shell python3 convert_hf_to_gguf.py /path/to/LFM2.5-Audio-1.5B/audio_detokenizer ``` * Update convert_hf_to_gguf.py Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com> * Formatting * Rework check for attention layers * Add LFM2 SWA model support * Address PR feedback * Set vocab to none * Move helper function definitions to cpp file --------- Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com> |
||
|
Ryan MangenoandSigbjørn Skjæret
|
c0d0430340 |
model : full modern bert support (#18330)
* full modern bert support * added gelu op in rank pooling for modern bert * still working on stuff, added mean calculation before classifier head * Update convert_hf_to_gguf.py Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com> * first layer is dense, as per modern bert research paper * Update src/llama-graph.cpp Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com> * fixed set input for mean pooling to check if pooling type is ranking since modern bert does mean & rank * Update src/llama-graph.cpp Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com> * Update convert_hf_to_gguf.py Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com> --------- Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com> |
||
|
ymckiandPiotr Wilkin
|
3688c4f504 |
Kimi-Linear support (backend agnostic + MLA KV cache) (#18755)
* kimi linear model implementation * kimi linear convert_hf_to_gguf * kimi linear constants.py tensor_mapping.py * Kimi Linear ggml.h * kimi linear ggml-cpu * Kimi Linear ggml-cuda * Kimi Linear ggml.c * kimi linear src/llama * remove "const int64_t n_seq_tokens = q->ne[2];" to get rid of unused variable warning * remove type mismatch warning * read MoE params * removed some hard coded code * removed all hard code * use DeepseekV2 tokenizer * removed unnecessary internal methods called by the old set_vocab of KimiLinear * rewrite get_vocab for KimiLinear. Removed all kda_scan code * removed all traces of kda_scan * reduce OP count by 1 due to removal of kda_scan * Move KIMI_LINEAR to llm_arch_is_hybrid to enable KV cache * set n_embd_head_k/v to ensure kv cache works * don't quantize conv1d of Kimi Linear * Kimi Linear backend agnostic * removed LOG_INFO * naive chunking form implemented * fixed some comments * add Kimi-K2 specific tokens to be recognized as EOG * build_kda_autoregressive is implemented to replace build_kda_recurrent for faster inference. sync'd to b7682 * replaced Akk and Aqk with mul_mat and clamp * no clamp version * Moved Aqk computation out of the loop * fixed typo and split wkv_b into wk_b and wv_b * MLA KV cache support * fix trailing spaces * moved const llama_model & model; around to follow qwen3next format and see if it cna pass the -Wunused-private-field error * fix trailing whitespace * removed traling whitespaces in empty line + make sure indentation is multiple of 4 * try to make lint happy * remove blank lines to make lint happy * removed at least blank line containing white space * fixed flake8 complaints locally * return ggml_tensor * pair in kda_autoregressive and kda_chunking as in ngxson's Qwen3Next improvement * removed Kimi-Linear specific change that causes failure at server-windows * removed private: from kimi_linear to make build checks happy * removed unnecessary ggml_cont before ggml_reshape * created static function causal_conv1d to abtract similar code for q/k/v * merged dt_bias to SSM_DT. Do -exp(log_A) in convert_hf_to_gguf.py. * reverted to original * fixed find_hparam calls. Fixed e_score_correction_bias to use bias instead of weight. Removed all ssm_conv bias terms. * remove DT_B from constants.py. remove one comment line in llama-model.cpp * new class llm_graph_input_mem_hybrid_k to get around the new MLA change. switch the concat order of ggml_concat calls in kimi-linear.cpp to accommodate MLA changes. Removed support for exp_probs_b.weight * remove ssm_o_norm_b * remove ssm_o_norm_b * changed hparams.kda_head_dim to hparams.n_embd_head_kda. added TODO comment for class llama_graph_mem_hybrid_k * removed all ggml_cont b4 ggml_reshape_4d * Whitespace * replaced all hparams.get with find_hparams * added new names for n_experts, n_experts_used and score_func in TextModel and removed their code in KimiLinear in convert_hf_to_gguf.py. Removed unnecessary ggml_cont and GGML_ASSERT in kimi-linear.cpp * use is_mla to switch between different mem_hybrid types * fixed logical errors in convert_hf_to_gguf.py pointed out by CISC * removed if else for required parameters kv_lora_rank and qk_rope_head_dim * add back ggml_cont for Vcur * minor changes * removed extra line in llama-vocab.cpp. Added back the comment in llama-graph.cpp * f16 gguf cannot run without context length * made a mistake of adding back n_ctx parsing --------- Co-authored-by: Piotr Wilkin (ilintar) <piotr.wilkin@syndatis.com> |
||
|
Georgi Gerganov
|
d9c6ce46f7 |
kv-cache : support V-less cache (#19067)
* kv-cache : support V-less cache * cuda : better check for V_is_K_view * cuda : improve V_is_K_view check * graph : add comments * hparams : refactor |
||
|
Georgi Gerganov
|
557515be1e |
graph : utilize ggml_build_forward_select() to avoid reallocations (#18898)
* graph : avoid branches between embedding and token inputs * models : make deepstack graphs (e.g. Qwen3 VL) have constant topology * ci : enable -DGGML_SCHED_NO_REALLOC=ON for server CI * cont : pad token embeddings to n_embd_inp |
||
|
Tarek DakhranandGeorgi Gerganov
|
ad8d85bd94 |
memory : add llama_memory_hybrid_iswa (#18601)
* memory : add llama_memory_hybrid_iswa * Update src/llama-memory-hybrid-iswa.cpp Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> |
||
|
|
d3dce4e0a5 |
sampling : add support for backend sampling (#17004)
* sampling : add support for backend sampling
This commit adds support for performing sampling operations on the
backend (e.g. GPU) as part of the model computation graph.
The motivation for this feature is to enable sampling to be performed
directly on the backend as part of the computation graph being executed,
allowing for some or all of the sampling to be done on the backend.
For example, the backend sampler chain might select/sample a token
directly in which case only the sampled token needs to be transferred
from device memory to host memory.
It is also possible for the backend samplers to perform filtering of
the logits, or compute and filter the probability distribution, in
which case only the filtered logits or probabilites need to be
transferred back to system memory for further processing by CPU
samplers.
Currently the backend sampling works in a similar manner to how
pooling works, it is a function that is called by build_graph and the
sampler operations become part of the models computation graph.
* llama-cli : add backend sampler configuration
* server : add backend sampling options/configuration
* webui : add backend sampling options
* ggml : add initial cumsum implementation for CUDA
* sampling : enable all backend sampler tests
This commit enables all exisiting backend sampler tests in the
test-backend-sampler. Previously, some tests were disabled because
there were missing ggml operation implementations.
* graph : do not include llama-model.h
* sampling : always expose sampled_ids
This commit precomputes and caches the full-vocab token id list in
llama_context's constructor, so llama_get_backend_sampled_token_ids_ith
always returns a valid pointer.
The motivation for this is that this enables both common/sampling.cpp
and src/llama-sampling.cpp can simplify their logic.
Not all backends samplers that process logits need to set the
sampled_tokens_id as they may not change the order of the logits, for
example the temperature sampler only scales the logits but does not
change their order. Simliar the logit bias sampler only adds bias to
specific token ids but does not change the order of the logits. In
these cases there will not be a device to host copy of the sampled
token ids, and this is the use case where having this precomputed
list is useful.
* sampling : ensure at most one output token per seq
This commit adds a check in the batch allocator to ensure that when
backend sampling is enabled, at most one output token is specified per
sequence.
* CUDA: Optimize argsort for gpu-based token sampling
Argsort is used for top-k currently. WE optimize argsort by 2 things:
1. Use `DeviceRadixSort` for single-row/sequence to parallelize it
across our SMs
2. Use `DeviceSegmentedSort` for multi-row/sequence as this is the
correct entrypoint (the function chooses different execution paths,
it contains `DeviceSegmentedRadixSort` as one of the paths and will
choose the best one according to heuristics.
https://nvidia.github.io/cccl/cub/api/structcub_1_1DeviceSegmentedSort.html#overview
Some perf numbers for a RTX PRO 6000:
On the kernel level, tested with
`GGML_CUDA_DISABLE_GRAPHS=1 ./test-backend-ops -o ARGSORT perf`
Before:
```
ARGSORT(type=f32,ne=[65000,16,1,1],order=0): 4130 runs - 359.24 us/run
ARGSORT(type=f32,ne=[200000,1,1,1],order=0): 8192 runs - 861.34 us/run
ARGSORT(type=f32,ne=[200000,16,1,1],order=0): 1343 runs - 1020.01 us/run
```
After:
```
ARGSORT(type=f32,ne=[65000,16,1,1],order=0): 4130 runs - 312.41 us/run
ARGSORT(type=f32,ne=[200000,1,1,1],order=0): 16384 runs - 63.48 us/run
ARGSORT(type=f32,ne=[200000,16,1,1],order=0): 1343 runs - 874.36 us/run
```
---
On the model level, tested with
`llama-cli -m gpt-oss-20b-mxfp4.gguf -n 200 -p "What is
the Capital of Sweden?" -no-cnv -fa 1 --backend-sampling`
Before:
```
llama_perf_sampler_print: sampling time = 0.25 ms / 207 runs ( 0.00 ms per token, 824701.20 tokens per second)
llama_perf_context_print: load time = 18215.58 ms
llama_perf_context_print: prompt eval time = 28.20 ms / 7 tokens ( 4.03 ms per token, 248.19 tokens per second)
llama_perf_context_print: eval time = 714.79 ms / 199 runs ( 3.59 ms per token, 278.40 tokens per second)
llama_perf_context_print: total time = 857.62 ms / 206 tokens
```
After
```
llama_perf_sampler_print: sampling time = 0.25 ms / 207 runs ( 0.00 ms per token, 828000.00 tokens per second)
llama_perf_context_print: load time = 18366.92 ms
llama_perf_context_print: prompt eval time = 35.92 ms / 7 tokens ( 5.13 ms per token, 194.87 tokens per second)
llama_perf_context_print: eval time = 532.79 ms / 199 runs ( 2.68 ms per token, 373.50 tokens per second)
llama_perf_context_print: total time = 683.65 ms / 206 tokens
```
* sampling : remove version from sampler chain
This commit removes the version field from the sampler chain and instead
used the sampler pointer itself for change detection.
* sampling : always populate logits for sampled probs
This commit updates common/sampler.cpp set_logits and
src/llama-sampling.cpp llama_sampler_sample to always populate the
logits field when backend sampled probabilities are available.
The motivation for this is that this ensure that CPU sampler always have
access to the logits values even when probabilites have been produced by
backend samplers.
* sampling : simplify backend sampling logic decode
This commit tries to simplify the backend sampling logic in
llama_context::decode.
* squash! sampling : simplify backend sampling logic decode
Fix condition to check if backend actually sampled tokens, not just that
backend samplers are available.
* common : fix regression caused by extra memory allocations during sampling
* squash! sampling : simplify backend sampling logic decode
The commit fixes a variable shadowing issue in the
`llama_context::decode` function which was introduced in a previous
refactoring.
* squash! common : fix regression caused by extra memory allocations during sampling
Apply the same changes to llama-sampling.cpp, llama_sampler_sample as
were applied in commit 38f408c25.
* sampling : introduce sampling_info struct
This commit introduces a sampling_info struct to encapsulate all
backend sampling related data within the llama_context class.
It also updates to use more descriptive names for sampled tokens and
candidates in the backend sampler ggml data structure.
* sampling : return early if backend sampling is disabled
* sampling : use pinned memory for backend sampling buffers
* common, tools : refactor model loading to support backend samplers
This commit refactors the model loading process in common/common.cpp
to enable backend sampler to be configure prior to the llama_context
creation.
The motivation for this change is that just being able to set/reset the
backend samplers after the llama_context has been created will cause a
resize to occur in llama_context::output_reserve which we want to avoid.
* sampling : add stride variable for clarity
* sampling: clarify candidate ids usage in comments
* sampling : fix copying both sampled tokens and logits/probs from backend
This commit fixes the issue where both sampled tokens and logits/probs
were not being copied correctly from the backend to the host when
multiple backend samplers were used.
A test for this scenario has also been added to ensure that both types
of data are copied correctly when different backend samplers are
employed.
* tests : cleanup test-backend-sampler.cpp
* common : remove build-info.cpp from commit [no ci]
This file was generated during the build process and should not be
included in previous commits.
* sampling : cleanup and clarify output_reserve
* sampling : remove redundant checks for stride and size [no ci]
* sampling : add debug log when backend sampler selects token
This commit adds a debug log statement in the llama_sampler_sample
to indicate when a backend sampler has selected a token for a given
index.
The modification helps in tracing the sampling process and understanding
the flow of control when backend samplers are used.
* examples : update batched to use backend sampling
This commit updates the batched example to demonstrate how to use
backend samplers.
* llama-cli : fix dangling reference to sampler config
* common : initialize backend samplers
* samplers : add missing cont
* sampling : add assertions for contiguous tensors in async copy functions
* examples : add info about hybrid sampling in batched [no ci]
* sampling : remove backend-dist option (wip)
This commit removes the `--backend-dist` option and instead uses the
configured --samplers chain to determine which samplers run on the
backend.
Backend sampling is still enabled using With `--backend_sampling`, and
the sampler chain, either explictly specified using `--samplers` or the
default, is automatically analyzed to determine which samplers can run
on the backend. The system finds the longest contiguous chain of
backend supported samplers from the start of the sampler sequence.
For example:
* If the chain is `top-k -> temperature -> top-p`, and both `top-k` and
`temperature` are backend-supported but `top-p` is not, then `top-k`
and `temperature` will run on the backend, while `top-p` and
subsequent samplers run on the CPU.
* If all configured samplers are supported, the final distribution
sampling will also happen on the backend, transferring only the
sampled token IDs back to the host.
* If the sampler chain starts with an unsupported sampler (e.g.,
`penalties`), all sampling runs on the CPU. Note that this is
currently the case with the default sampler so to use backend sampling
it is required to specify a sampler chain. See below for an example.
The following shows how llama-cli can be run with backend sampling:
```console
$ llama-cli -m models/Qwen2.5-VL-3B-Instruct-Q8_0.gguf \
--prompt 'What is the capital of Sweden?' \
-n 20 \
-no-cnv \
--verbose-prompt \
-ngl 40 \
--backend-sampling \
--samplers 'top_k;temperature'
```
In this case the all sampling will happen on the backend since both
`top_k` and `temperature` are supported backend samplers.
To enable a partial backend sampling (hybrid sampling), for example
running `top_k` and `temperature` on the backend and `typ_p` on the CPU
the following sampler chain could be specified:
```console
$ llama-cli -m models/Qwen2.5-VL-3B-Instruct-Q8_0.gguf \
--prompt 'What is the capital of Sweden?' \
-n 20 \
-no-cnv \
--verbose-prompt \
-ngl 40 \
--backend-sampling \
--samplers 'top_k;temperature;top_p'
```
If this looks good then I'll follow up with updates the llama-cli and
llama-server documentation to reflect these changes.
* CUDA: Add top-k implementation
* sampling : add min-p backend sampler
* Use `FetchContent` over CPM as it's bundled with CMake
Thanks @ggerganov for the suggestion
* common : add get_active_samplers function to check enabled samplers
This commit adds a function to check if a sampler is actually enabled,
meaning that it does not have values that disables its effect. This is
then used by the backend samplers initialization to avoid considering
samplers that are not enabled when determining the split point between
them.
The motivation for this is that this allows the default sampler chain
for `--samplers` to be used and any sampler that is not enabled will not
cause the backend samplers to be skipped.
For example, before this change if the penalties sampler was included in
the samplers list but had default values that disable it, it would cause
the backend samplers to be skipped entirely.
This commit also contains some refactoring to remove some code
duplication.
* cuda : fix editorconfig-checker warning
* sampling : use argmax for min-p sampling
* sampling : fix temperature check to allow zero temperature
This commit modifies the temperature sampling check to allow a
temperature value of zero. Previously, the check only allowed
positive temperature values, which excluded the valid case of
zero temperature.
The motivation for this is to enable a zero temperature setting which is
also currently causing the following test to fail:
```console
(venv) $ cd tools/server/tests
(venv) $ ./tests.sh unit/test_basic.py::test_load_split_model
```
* cuda : fix top-k compilation when CUB is unavailable
This commit adds a macro guard around argsort_f32_i32_cuda_cub usage
in the top-k fallback path, falling back to bitonic sort when
GGML_CUDA_USE_CUB is not defined.
The motivation for this is that some environments like AMD HIP
do not have CUB available, causing compilation failure.
Refs: https://github.com/ggml-org/llama.cpp/actions/runs/19728226426/job/56523606840#step:6:208
* sampling : add comments about backend sampler [no ci]
This commit adds a comment to llama_context's constructor explaining why
backend samplers are initialized early in the process.
* sampling : remove backend sampling chain from common_sampler
This commit removes the backend sampling chain from the common_sampler
structure and related functions.
The motivation for this change is that the backend samplers are not
currently set on the context, and if they are they would cause the
a graph reallocation to occur. Instead, the intialization is handled
like it currently is by llama_context's constructor.
* Fix top-k comp & behavior for non-CUB path
Some changes were made in 5ea3be265ba6f8916daf52e19e3fb8efe9a03637
which were incomplete. In the case of non-CUB, bitonic sort and its
limitations of ncols < 1024 have to apply, similar to argsort.cu
* sampling : support intermixed backend/cpu samplers
This commit updates the backend sampling implementation to support
intermixed usage of backend and CPU samplers within the same batch.
The initial implementation was developed as an all-or-nothing solution:
either perform backend sampling for the entire batch, or perform CPU
sampling for the entire batch.
The motivation for this change is to support batches with mixed
sequences. For example, we may have a backend sampler configured for
sequence 0, while sequence 1 in the same batch uses CPU sampling. This
was not supported in the initial implementation.
This issue manifested in llama-server with the webui: decoding with
backend samplers would work initially, but after changing to CPU
sampling, a slot (sequence) could still be using a backend sampler.
This meant that logits in output_reserve would not be allocated,
resulting in an error.
The solution in this commit inspects the batch to determine which
sampling modes are needed and allocates buffers accordingly. However,
there is a known inefficiency: when we have intermixed backend/CPU
samplers in the same batch, we currently copy all logits to the host,
even for sequences using backend samplers.
Added test_backend_cpu_mixed_batch to verify correct behavior with
mixed backend/CPU samplers in a single batch, including dynamic
sampler switching between decode calls.
* squash! sampling : support intermixed backend/cpu samplers
Add check that logits is not null which is can happen for embeddings.
* squash! sampling : support intermixed backend/cpu samplers
Fix llama-save-load-state which currently fails by handling the case
when batch.logits is nullptr (like when loading state) by allocating
space for all outputs as CPU logits.
* refactor : simplify and improve memory management
* Add initial version for top-p sampling
As we only support static graphs for the time and we don't know the size
of the output of top-p, we have to do value-scaling same as for min-p
operator.
Further improvements can be applied to the unit-test (i.e. check for
equivalence of top_p happening on backend with top_p happening on cpu)
and also by constructing candidates and sorting those as opposed to
reversing the sort of the logits (this would be arange +
get_rows instead of argsort + get_rows)
* sampling : use logits directly for min-p filtering
* sampling : simplify
* llama : simplify
* llama : cleanup + naming
* llama : call backend_init once
* llama : reserve graphs with samplers
* llama : naming
* cont : naming
* sampling : lower log level for output buffer reallocations [no ci]
This commit changes the logging level for output buffer reallocations
in the llama_context::output_reserve function from INFO to DEBUG.
The motivation for this is that it currently logs to info and when
enabling verbose logging for llama-cli this will get mixed with the
output, for example:
```console
What is the capital of Sweden?output_reserve: reallocating output buffer from size 0.58 MiB to 1.74 MiB
1. Stockholm
2\. Helsinki
Based are the options
1. Stockholm
Explanation: Stockholm is the capital of
...
```
* Fix backend_top_p_sampler
softmax(softmax) will return uniform distribution, so we should not
return the softmax but the logits instead.
* Factor out `ggml_sort` into its own function
* Make backend's top_p sampler inclusive
In addition to match the algorithm proposed in the original
[paper](https://arxiv.org/abs/1904.09751), this resolves the edge-case
where `max_p is > top_p` for a single logit, where the mask would
otherwise be empty (and we thus sample from the whole vocabulary with
equal likelihood)
* common : simplify sampler chain initialization
* sampling : do not create empty samplers
* sampling : fix top_p empty condition
* examples : remove outdated backend sampling section
This commit removes the outdated section about using backend samplers
from the README.md file in the examples/batched.
* sampling : fix backend temp sampler for zero temperature
This commit fixes the implementation of the temperature-based sampler
for the case when the temperature is set to zero. This now correctly
selects the most probable token by masking out all other tokens in the
logits.
* CUDA: Move cccl fetch to after cuda has been enabled in CMakeLists.txt
This will allow cccl to set build flags for the CUDA compiler, required
e.g. for MSVC compat, see also
https://github.com/NVIDIA/cccl/pull/6791
* CUDA: Use standard-compliant preprocessor for MSVC builds
Workarounds of https://github.com/NVIDIA/cccl/pull/6791 will not be
backported to CCCL 3.2, only the diagnostics/error messages will:
https://github.com/NVIDIA/cccl/pull/6827
* CUDA: Update CCCL's rc candidate
* squash! sampling : fix backend temp sampler for zero temperature
This modifies the parent commit to simply return the most probably token
instead of masking the logits.
* sampling : implement temp_ext_backend sampling
This commit implements the apply function for the extended temperature
sampling.
* sampling : minor cleanup
* sampling : stop short if backend sampler sampled a token
This commit modifies the graph building logic to immediately continue
when a token has already been sampled by the backend sampler.
It also updates the test for backend temporary sampling to include
top-k and distribution samplers in the chain to verify that they are not
producing any logits (they are not run).
* Revert "sampling : stop short if backend sampler sampled a token"
This reverts commit 87b2719eca55b30afff600fc7f61c6cce9452cbf.
* sampling : fix backend temp sampling to use logits masking
* sampling : simplify temp sampling
* sampling : remove redundant calls to ggml_build_forward_expand
* sampling : check backend support during init
* cont : keep backend sampling disabled for now
* sampling : fix outputs and device checks
* sampling : fix candidates logic
* Add perf-tests for CUMSUM
* Readd `cub::DeviceScan::InclusiveSum`-based CumSum
For single rows and large columns doing a for-loop over the function
`cub::DeviceScan::InclusiveSum` offered by CUB outperforms the
`cumsum_cub_kernel` where `cub::BlockScan` is used.
Numbers before this change
Backend 1/3: CUDA0
Device description: NVIDIA RTX 6000 Ada Generation
Device memory: 48510 MB (48039 MB free)
CUMSUM(type=f32,ne=[128,128,4,4]): 311258 runs - 3.26 us/run - 2048 kB/run - 599.76 GB/s
CUMSUM(type=f32,ne=[2048,16,5,4]): 229390 runs - 4.40 us/run - 5120 kB/run - 1110.23 GB/s
CUMSUM(type=f32,ne=[20000,10,4,1]): 37583 runs - 29.63 us/run - 6250 kB/run - 201.18 GB/s
CUMSUM(type=f32,ne=[128,1,1,1]): 892819 runs - 1.12 us/run - 1 kB/run - 0.85 GB/s
CUMSUM(type=f32,ne=[1024,1,1,1]): 450505 runs - 2.25 us/run - 8 kB/run - 3.39 GB/s
CUMSUM(type=f32,ne=[4096,1,1,1]): 155629 runs - 6.61 us/run - 32 kB/run - 4.62 GB/s
CUMSUM(type=f32,ne=[8192,1,1,1]): 81910 runs - 12.60 us/run - 64 kB/run - 4.85 GB/s
CUMSUM(type=f32,ne=[16384,1,1,1]): 49146 runs - 23.99 us/run - 128 kB/run - 5.09 GB/s
CUMSUM(type=f32,ne=[32768,1,1,1]): 24573 runs - 47.10 us/run - 256 kB/run - 5.18 GB/s
CUMSUM(type=f32,ne=[65536,1,1,1]): 16382 runs - 93.57 us/run - 512 kB/run - 5.22 GB/s
CUMSUM(type=f32,ne=[131072,1,1,1]): 8191 runs - 184.79 us/run - 1024 kB/run - 5.29 GB/s
CUMSUM(type=f32,ne=[200000,1,1,1]): 8191 runs - 280.43 us/run - 1562 kB/run - 5.31 GB/s
CUMSUM(type=f32,ne=[2000000,1,1,1]): 2148 runs - 2771.23 us/run - 15625 kB/run - 5.38 GB/s
CUMSUM(type=f32,ne=[128,4,1,1]): 458696 runs - 2.21 us/run - 4 kB/run - 1.73 GB/s
CUMSUM(type=f32,ne=[1024,4,1,1]): 360404 runs - 2.82 us/run - 32 kB/run - 10.83 GB/s
CUMSUM(type=f32,ne=[4096,4,1,1]): 147438 runs - 7.12 us/run - 128 kB/run - 17.15 GB/s
CUMSUM(type=f32,ne=[8192,4,1,1]): 81910 runs - 12.90 us/run - 256 kB/run - 18.92 GB/s
CUMSUM(type=f32,ne=[16384,4,1,1]): 49146 runs - 24.32 us/run - 512 kB/run - 20.08 GB/s
CUMSUM(type=f32,ne=[32768,4,1,1]): 24573 runs - 47.28 us/run - 1024 kB/run - 20.66 GB/s
CUMSUM(type=f32,ne=[65536,4,1,1]): 16382 runs - 93.21 us/run - 2048 kB/run - 20.96 GB/s
CUMSUM(type=f32,ne=[131072,4,1,1]): 8191 runs - 185.04 us/run - 4096 kB/run - 21.11 GB/s
CUMSUM(type=f32,ne=[200000,4,1,1]): 5369 runs - 282.08 us/run - 6250 kB/run - 21.13 GB/s
CUMSUM(type=f32,ne=[2000000,4,1,1]): 537 runs - 2806.46 us/run - 62500 kB/run - 21.26 GB/s
CUMSUM(type=f32,ne=[128,8,1,1]): 458696 runs - 2.20 us/run - 8 kB/run - 3.47 GB/s
CUMSUM(type=f32,ne=[1024,8,1,1]): 360404 runs - 2.82 us/run - 64 kB/run - 21.66 GB/s
CUMSUM(type=f32,ne=[4096,8,1,1]): 147438 runs - 7.12 us/run - 256 kB/run - 34.28 GB/s
CUMSUM(type=f32,ne=[8192,8,1,1]): 81910 runs - 12.90 us/run - 512 kB/run - 37.84 GB/s
CUMSUM(type=f32,ne=[16384,8,1,1]): 49146 runs - 24.32 us/run - 1024 kB/run - 40.15 GB/s
CUMSUM(type=f32,ne=[32768,8,1,1]): 24573 runs - 47.28 us/run - 2048 kB/run - 41.31 GB/s
CUMSUM(type=f32,ne=[65536,8,1,1]): 16382 runs - 93.20 us/run - 4096 kB/run - 41.92 GB/s
CUMSUM(type=f32,ne=[131072,8,1,1]): 8194 runs - 185.05 us/run - 8192 kB/run - 42.22 GB/s
CUMSUM(type=f32,ne=[200000,8,1,1]): 5370 runs - 282.15 us/run - 12500 kB/run - 42.26 GB/s
CUMSUM(type=f32,ne=[2000000,8,1,1]): 269 runs - 4067.61 us/run - 125000 kB/run - 29.36 GB/s
CUMSUM(type=f32,ne=[128,16,1,1]): 303067 runs - 3.32 us/run - 16 kB/run - 4.60 GB/s
CUMSUM(type=f32,ne=[1024,16,1,1]): 303067 runs - 3.32 us/run - 128 kB/run - 36.76 GB/s
CUMSUM(type=f32,ne=[4096,16,1,1]): 147438 runs - 7.17 us/run - 512 kB/run - 68.13 GB/s
CUMSUM(type=f32,ne=[8192,16,1,1]): 81910 runs - 12.90 us/run - 1024 kB/run - 75.68 GB/s
CUMSUM(type=f32,ne=[16384,16,1,1]): 49146 runs - 24.33 us/run - 2048 kB/run - 80.28 GB/s
CUMSUM(type=f32,ne=[32768,16,1,1]): 24573 runs - 47.30 us/run - 4096 kB/run - 82.59 GB/s
CUMSUM(type=f32,ne=[65536,16,1,1]): 12291 runs - 93.24 us/run - 8192 kB/run - 83.80 GB/s
CUMSUM(type=f32,ne=[131072,16,1,1]): 6147 runs - 185.07 us/run - 16384 kB/run - 84.45 GB/s
CUMSUM(type=f32,ne=[200000,16,1,1]): 4029 runs - 282.40 us/run - 25000 kB/run - 84.46 GB/s
CUMSUM(type=f32,ne=[2000000,16,1,1]): 270 runs - 4118.40 us/run - 250000 kB/run - 58.11 GB/s
Backend CUDA0: OK
Backend 2/3: CUDA1
Device description: NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition
Device memory: 97250 MB (96677 MB free)
CUMSUM(type=f32,ne=[128,128,4,4]): 368595 runs - 2.73 us/run - 2048 kB/run - 715.83 GB/s
CUMSUM(type=f32,ne=[2048,16,5,4]): 216282 runs - 4.72 us/run - 5120 kB/run - 1035.32 GB/s
CUMSUM(type=f32,ne=[20000,10,4,1]): 32214 runs - 34.33 us/run - 6250 kB/run - 173.64 GB/s
CUMSUM(type=f32,ne=[128,1,1,1]): 810909 runs - 1.24 us/run - 1 kB/run - 0.77 GB/s
CUMSUM(type=f32,ne=[1024,1,1,1]): 401359 runs - 2.52 us/run - 8 kB/run - 3.03 GB/s
CUMSUM(type=f32,ne=[4096,1,1,1]): 139247 runs - 7.44 us/run - 32 kB/run - 4.10 GB/s
CUMSUM(type=f32,ne=[8192,1,1,1]): 73719 runs - 14.27 us/run - 64 kB/run - 4.28 GB/s
CUMSUM(type=f32,ne=[16384,1,1,1]): 40955 runs - 27.24 us/run - 128 kB/run - 4.48 GB/s
CUMSUM(type=f32,ne=[32768,1,1,1]): 24573 runs - 53.46 us/run - 256 kB/run - 4.57 GB/s
CUMSUM(type=f32,ne=[65536,1,1,1]): 16382 runs - 105.29 us/run - 512 kB/run - 4.64 GB/s
CUMSUM(type=f32,ne=[131072,1,1,1]): 8191 runs - 210.15 us/run - 1024 kB/run - 4.65 GB/s
CUMSUM(type=f32,ne=[200000,1,1,1]): 8191 runs - 318.22 us/run - 1562 kB/run - 4.68 GB/s
CUMSUM(type=f32,ne=[2000000,1,1,1]): 2148 runs - 3142.23 us/run - 15625 kB/run - 4.74 GB/s
CUMSUM(type=f32,ne=[128,4,1,1]): 303067 runs - 3.34 us/run - 4 kB/run - 1.14 GB/s
CUMSUM(type=f32,ne=[1024,4,1,1]): 253921 runs - 4.03 us/run - 32 kB/run - 7.58 GB/s
CUMSUM(type=f32,ne=[4096,4,1,1]): 122865 runs - 8.20 us/run - 128 kB/run - 14.89 GB/s
CUMSUM(type=f32,ne=[8192,4,1,1]): 73719 runs - 14.96 us/run - 256 kB/run - 16.32 GB/s
CUMSUM(type=f32,ne=[16384,4,1,1]): 40955 runs - 28.66 us/run - 512 kB/run - 17.04 GB/s
CUMSUM(type=f32,ne=[32768,4,1,1]): 24573 runs - 54.21 us/run - 1024 kB/run - 18.01 GB/s
CUMSUM(type=f32,ne=[65536,4,1,1]): 16382 runs - 106.49 us/run - 2048 kB/run - 18.34 GB/s
CUMSUM(type=f32,ne=[131072,4,1,1]): 8191 runs - 210.88 us/run - 4096 kB/run - 18.52 GB/s
CUMSUM(type=f32,ne=[200000,4,1,1]): 5369 runs - 321.77 us/run - 6250 kB/run - 18.53 GB/s
CUMSUM(type=f32,ne=[2000000,4,1,1]): 537 runs - 3191.79 us/run - 62500 kB/run - 18.69 GB/s
CUMSUM(type=f32,ne=[128,8,1,1]): 376786 runs - 2.67 us/run - 8 kB/run - 2.86 GB/s
CUMSUM(type=f32,ne=[1024,8,1,1]): 245730 runs - 4.10 us/run - 64 kB/run - 14.90 GB/s
CUMSUM(type=f32,ne=[4096,8,1,1]): 122865 runs - 8.20 us/run - 256 kB/run - 29.79 GB/s
CUMSUM(type=f32,ne=[8192,8,1,1]): 65528 runs - 16.38 us/run - 512 kB/run - 29.82 GB/s
CUMSUM(type=f32,ne=[16384,8,1,1]): 40955 runs - 28.69 us/run - 1024 kB/run - 34.04 GB/s
CUMSUM(type=f32,ne=[32768,8,1,1]): 24573 runs - 55.28 us/run - 2048 kB/run - 35.33 GB/s
CUMSUM(type=f32,ne=[65536,8,1,1]): 16382 runs - 108.50 us/run - 4096 kB/run - 36.00 GB/s
CUMSUM(type=f32,ne=[131072,8,1,1]): 8194 runs - 213.75 us/run - 8192 kB/run - 36.55 GB/s
CUMSUM(type=f32,ne=[200000,8,1,1]): 5370 runs - 326.31 us/run - 12500 kB/run - 36.54 GB/s
CUMSUM(type=f32,ne=[2000000,8,1,1]): 538 runs - 3252.68 us/run - 125000 kB/run - 36.72 GB/s
CUMSUM(type=f32,ne=[128,16,1,1]): 303067 runs - 3.32 us/run - 16 kB/run - 4.60 GB/s
CUMSUM(type=f32,ne=[1024,16,1,1]): 253921 runs - 4.06 us/run - 128 kB/run - 30.09 GB/s
CUMSUM(type=f32,ne=[4096,16,1,1]): 122865 runs - 8.20 us/run - 512 kB/run - 59.57 GB/s
CUMSUM(type=f32,ne=[8192,16,1,1]): 65528 runs - 16.38 us/run - 1024 kB/run - 59.63 GB/s
CUMSUM(type=f32,ne=[16384,16,1,1]): 40955 runs - 28.69 us/run - 2048 kB/run - 68.09 GB/s
CUMSUM(type=f32,ne=[32768,16,1,1]): 24573 runs - 55.28 us/run - 4096 kB/run - 70.67 GB/s
CUMSUM(type=f32,ne=[65536,16,1,1]): 12291 runs - 108.50 us/run - 8192 kB/run - 72.02 GB/s
CUMSUM(type=f32,ne=[131072,16,1,1]): 6147 runs - 213.60 us/run - 16384 kB/run - 73.17 GB/s
CUMSUM(type=f32,ne=[200000,16,1,1]): 4029 runs - 326.04 us/run - 25000 kB/run - 73.15 GB/s
CUMSUM(type=f32,ne=[2000000,16,1,1]): 270 runs - 5458.69 us/run - 250000 kB/run - 43.84 GB/s
----
Numbers after:
Backend 1/3: CUDA0
Device description: NVIDIA RTX 6000 Ada Generation
Device memory: 48510 MB (48039 MB free)
CUMSUM(type=f32,ne=[128,128,4,4]): 311258 runs - 3.25 us/run - 2048 kB/run - 601.62 GB/s
CUMSUM(type=f32,ne=[2048,16,5,4]): 229390 runs - 4.40 us/run - 5120 kB/run - 1110.14 GB/s
CUMSUM(type=f32,ne=[20000,10,4,1]): 37583 runs - 29.67 us/run - 6250 kB/run - 200.89 GB/s
CUMSUM(type=f32,ne=[128,1,1,1]): 892819 runs - 1.12 us/run - 1 kB/run - 0.85 GB/s
CUMSUM(type=f32,ne=[1024,1,1,1]): 458696 runs - 2.21 us/run - 8 kB/run - 3.45 GB/s
CUMSUM(type=f32,ne=[4096,1,1,1]): 376786 runs - 2.66 us/run - 32 kB/run - 11.46 GB/s
CUMSUM(type=f32,ne=[8192,1,1,1]): 393168 runs - 2.59 us/run - 64 kB/run - 23.57 GB/s
CUMSUM(type=f32,ne=[16384,1,1,1]): 393168 runs - 2.59 us/run - 128 kB/run - 47.15 GB/s
CUMSUM(type=f32,ne=[32768,1,1,1]): 376786 runs - 2.69 us/run - 256 kB/run - 90.69 GB/s
CUMSUM(type=f32,ne=[65536,1,1,1]): 327640 runs - 3.06 us/run - 512 kB/run - 159.65 GB/s
CUMSUM(type=f32,ne=[131072,1,1,1]): 311258 runs - 3.28 us/run - 1024 kB/run - 297.77 GB/s
CUMSUM(type=f32,ne=[200000,1,1,1]): 270303 runs - 3.74 us/run - 1562 kB/run - 398.14 GB/s
CUMSUM(type=f32,ne=[2000000,1,1,1]): 137472 runs - 7.35 us/run - 15625 kB/run - 2026.94 GB/s
CUMSUM(type=f32,ne=[128,4,1,1]): 876437 runs - 1.14 us/run - 4 kB/run - 3.33 GB/s
CUMSUM(type=f32,ne=[1024,4,1,1]): 442314 runs - 2.28 us/run - 32 kB/run - 13.39 GB/s
CUMSUM(type=f32,ne=[4096,4,1,1]): 155629 runs - 6.69 us/run - 128 kB/run - 18.24 GB/s
CUMSUM(type=f32,ne=[8192,4,1,1]): 81910 runs - 12.53 us/run - 256 kB/run - 19.49 GB/s
CUMSUM(type=f32,ne=[16384,4,1,1]): 49146 runs - 24.18 us/run - 512 kB/run - 20.20 GB/s
CUMSUM(type=f32,ne=[32768,4,1,1]): 65528 runs - 15.34 us/run - 1024 kB/run - 63.66 GB/s
CUMSUM(type=f32,ne=[65536,4,1,1]): 73719 runs - 14.76 us/run - 2048 kB/run - 132.35 GB/s
CUMSUM(type=f32,ne=[131072,4,1,1]): 65528 runs - 16.01 us/run - 4096 kB/run - 244.07 GB/s
CUMSUM(type=f32,ne=[200000,4,1,1]): 64428 runs - 16.51 us/run - 6250 kB/run - 360.97 GB/s
CUMSUM(type=f32,ne=[2000000,4,1,1]): 33831 runs - 29.59 us/run - 62500 kB/run - 2016.08 GB/s
CUMSUM(type=f32,ne=[128,8,1,1]): 868246 runs - 1.16 us/run - 8 kB/run - 6.59 GB/s
CUMSUM(type=f32,ne=[1024,8,1,1]): 442314 runs - 2.28 us/run - 64 kB/run - 26.76 GB/s
CUMSUM(type=f32,ne=[4096,8,1,1]): 155629 runs - 6.69 us/run - 256 kB/run - 36.48 GB/s
CUMSUM(type=f32,ne=[8192,8,1,1]): 81910 runs - 12.53 us/run - 512 kB/run - 38.97 GB/s
CUMSUM(type=f32,ne=[16384,8,1,1]): 49146 runs - 24.17 us/run - 1024 kB/run - 40.41 GB/s
CUMSUM(type=f32,ne=[32768,8,1,1]): 24573 runs - 47.53 us/run - 2048 kB/run - 41.10 GB/s
CUMSUM(type=f32,ne=[65536,8,1,1]): 16382 runs - 61.25 us/run - 4096 kB/run - 63.77 GB/s
CUMSUM(type=f32,ne=[131072,8,1,1]): 32776 runs - 31.79 us/run - 8192 kB/run - 245.82 GB/s
CUMSUM(type=f32,ne=[200000,8,1,1]): 32220 runs - 32.90 us/run - 12500 kB/run - 362.35 GB/s
CUMSUM(type=f32,ne=[2000000,8,1,1]): 6725 runs - 151.99 us/run - 125000 kB/run - 785.77 GB/s
CUMSUM(type=f32,ne=[128,16,1,1]): 851864 runs - 1.18 us/run - 16 kB/run - 12.97 GB/s
CUMSUM(type=f32,ne=[1024,16,1,1]): 442314 runs - 2.30 us/run - 128 kB/run - 53.13 GB/s
CUMSUM(type=f32,ne=[4096,16,1,1]): 155629 runs - 6.68 us/run - 512 kB/run - 73.13 GB/s
CUMSUM(type=f32,ne=[8192,16,1,1]): 81910 runs - 12.68 us/run - 1024 kB/run - 77.00 GB/s
CUMSUM(type=f32,ne=[16384,16,1,1]): 40955 runs - 24.56 us/run - 2048 kB/run - 79.53 GB/s
CUMSUM(type=f32,ne=[32768,16,1,1]): 24573 runs - 47.52 us/run - 4096 kB/run - 82.21 GB/s
CUMSUM(type=f32,ne=[65536,16,1,1]): 12291 runs - 93.44 us/run - 8192 kB/run - 83.62 GB/s
CUMSUM(type=f32,ne=[131072,16,1,1]): 16392 runs - 63.36 us/run - 16384 kB/run - 246.68 GB/s
CUMSUM(type=f32,ne=[200000,16,1,1]): 16116 runs - 65.25 us/run - 25000 kB/run - 365.53 GB/s
CUMSUM(type=f32,ne=[2000000,16,1,1]): 3375 runs - 304.46 us/run - 250000 kB/run - 785.98 GB/s
Backend CUDA0: OK
Backend 2/3: CUDA1
Device description: NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition
Device memory: 97250 MB (96677 MB free)
CUMSUM(type=f32,ne=[128,128,4,4]): 376786 runs - 2.69 us/run - 2048 kB/run - 727.04 GB/s
CUMSUM(type=f32,ne=[2048,16,5,4]): 216282 runs - 4.64 us/run - 5120 kB/run - 1053.30 GB/s
CUMSUM(type=f32,ne=[20000,10,4,1]): 32214 runs - 34.21 us/run - 6250 kB/run - 174.27 GB/s
CUMSUM(type=f32,ne=[128,1,1,1]): 819100 runs - 1.22 us/run - 1 kB/run - 0.78 GB/s
CUMSUM(type=f32,ne=[1024,1,1,1]): 409550 runs - 2.47 us/run - 8 kB/run - 3.09 GB/s
CUMSUM(type=f32,ne=[4096,1,1,1]): 303067 runs - 3.31 us/run - 32 kB/run - 9.21 GB/s
CUMSUM(type=f32,ne=[8192,1,1,1]): 237539 runs - 4.33 us/run - 64 kB/run - 14.08 GB/s
CUMSUM(type=f32,ne=[16384,1,1,1]): 237539 runs - 4.33 us/run - 128 kB/run - 28.17 GB/s
CUMSUM(type=f32,ne=[32768,1,1,1]): 188393 runs - 5.37 us/run - 256 kB/run - 45.47 GB/s
CUMSUM(type=f32,ne=[65536,1,1,1]): 188393 runs - 5.41 us/run - 512 kB/run - 90.20 GB/s
CUMSUM(type=f32,ne=[131072,1,1,1]): 188393 runs - 5.41 us/run - 1024 kB/run - 180.41 GB/s
CUMSUM(type=f32,ne=[200000,1,1,1]): 188393 runs - 5.41 us/run - 1562 kB/run - 275.27 GB/s
CUMSUM(type=f32,ne=[2000000,1,1,1]): 128880 runs - 7.76 us/run - 15625 kB/run - 1920.33 GB/s
CUMSUM(type=f32,ne=[128,4,1,1]): 802718 runs - 1.26 us/run - 4 kB/run - 3.03 GB/s
CUMSUM(type=f32,ne=[1024,4,1,1]): 401359 runs - 2.51 us/run - 32 kB/run - 12.18 GB/s
CUMSUM(type=f32,ne=[4096,4,1,1]): 139247 runs - 7.51 us/run - 128 kB/run - 16.26 GB/s
CUMSUM(type=f32,ne=[8192,4,1,1]): 73719 runs - 14.17 us/run - 256 kB/run - 17.23 GB/s
CUMSUM(type=f32,ne=[16384,4,1,1]): 40955 runs - 27.37 us/run - 512 kB/run - 17.84 GB/s
CUMSUM(type=f32,ne=[32768,4,1,1]): 40955 runs - 26.33 us/run - 1024 kB/run - 37.10 GB/s
CUMSUM(type=f32,ne=[65536,4,1,1]): 40955 runs - 26.19 us/run - 2048 kB/run - 74.59 GB/s
CUMSUM(type=f32,ne=[131072,4,1,1]): 40955 runs - 26.35 us/run - 4096 kB/run - 148.26 GB/s
CUMSUM(type=f32,ne=[200000,4,1,1]): 42952 runs - 24.18 us/run - 6250 kB/run - 246.51 GB/s
CUMSUM(type=f32,ne=[2000000,4,1,1]): 32757 runs - 31.01 us/run - 62500 kB/run - 1923.68 GB/s
CUMSUM(type=f32,ne=[128,8,1,1]): 786336 runs - 1.28 us/run - 8 kB/run - 5.95 GB/s
CUMSUM(type=f32,ne=[1024,8,1,1]): 393168 runs - 2.57 us/run - 64 kB/run - 23.73 GB/s
CUMSUM(type=f32,ne=[4096,8,1,1]): 131056 runs - 7.67 us/run - 256 kB/run - 31.82 GB/s
CUMSUM(type=f32,ne=[8192,8,1,1]): 73719 runs - 14.43 us/run - 512 kB/run - 33.84 GB/s
CUMSUM(type=f32,ne=[16384,8,1,1]): 40955 runs - 27.90 us/run - 1024 kB/run - 35.01 GB/s
CUMSUM(type=f32,ne=[32768,8,1,1]): 24573 runs - 54.63 us/run - 2048 kB/run - 35.75 GB/s
CUMSUM(type=f32,ne=[65536,8,1,1]): 16382 runs - 72.24 us/run - 4096 kB/run - 54.08 GB/s
CUMSUM(type=f32,ne=[131072,8,1,1]): 20485 runs - 52.66 us/run - 8192 kB/run - 148.37 GB/s
CUMSUM(type=f32,ne=[200000,8,1,1]): 21480 runs - 48.00 us/run - 12500 kB/run - 248.42 GB/s
CUMSUM(type=f32,ne=[2000000,8,1,1]): 16140 runs - 61.99 us/run - 125000 kB/run - 1926.51 GB/s
CUMSUM(type=f32,ne=[128,16,1,1]): 786336 runs - 1.28 us/run - 16 kB/run - 11.90 GB/s
CUMSUM(type=f32,ne=[1024,16,1,1]): 393168 runs - 2.57 us/run - 128 kB/run - 47.57 GB/s
CUMSUM(type=f32,ne=[4096,16,1,1]): 131056 runs - 7.65 us/run - 512 kB/run - 63.83 GB/s
CUMSUM(type=f32,ne=[8192,16,1,1]): 73719 runs - 14.42 us/run - 1024 kB/run - 67.74 GB/s
CUMSUM(type=f32,ne=[16384,16,1,1]): 40955 runs - 27.87 us/run - 2048 kB/run - 70.09 GB/s
CUMSUM(type=f32,ne=[32768,16,1,1]): 24573 runs - 54.54 us/run - 4096 kB/run - 71.63 GB/s
CUMSUM(type=f32,ne=[65536,16,1,1]): 12291 runs - 107.53 us/run - 8192 kB/run - 72.66 GB/s
CUMSUM(type=f32,ne=[131072,16,1,1]): 10245 runs - 105.10 us/run - 16384 kB/run - 148.70 GB/s
CUMSUM(type=f32,ne=[200000,16,1,1]): 10744 runs - 95.36 us/run - 25000 kB/run - 250.11 GB/s
CUMSUM(type=f32,ne=[2000000,16,1,1]): 5400 runs - 186.97 us/run - 250000 kB/run - 1279.90 GB/s
* sampling : expand support (wip)
* tests : fix memory leaks
* cont : fixes
* tests : check temp back to 0.0
* sampling : fix top-p
* sampling : handle n_probs case
* server : handle unsupported cases
* metal : print node names for debugging
* ggml : remove redundant src in ggml_cast
* ggml-alloc : fix reuse-parent logic for misaligned sizes
* Revert "ggml : remove redundant src in ggml_cast"
This reverts commit 62d1b0082dbad699fbeea85a096bc334e3c1c0e6.
* CUDA: Add Cooperative-Groups-based parallelization of ncols in softmax
Old implementation parallelizes rows across SMs, which does not fit the
needs of backend-sampling (where we have ncols >> nrows and thus want to
parallelize ncols across SMs)
* Add TODOs to and adjust heuristics of row-wise soft_max in CUDA
Heuristics were selected based on the following numbers:
```
-- Before
Backend 1/2: CUDA0
Device description: NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition
Device memory: 97250 MB (96691 MB free)
SOFT_MAX(type=f32,ne=[4096,4096,5,1],mask=0,sinks=0,m_prec=f32,nr23=[1,1],scale=1.000000,max_bias=0.000000,inplace=0): 2236 runs - 450.34 us/run - 655360 kB/run - 1401.20 GB/s
SOFT_MAX(type=f32,ne=[12888,256,5,1],mask=0,sinks=0,m_prec=f32,nr23=[1,1],scale=1.000000,max_bias=0.000000,inplace=0): 17748 runs - 56.80 us/run - 128880 kB/run - 2168.19 GB/s
SOFT_MAX(type=f32,ne=[77,4096,5,1],mask=0,sinks=0,m_prec=f32,nr23=[1,1],scale=1.000000,max_bias=0.000000,inplace=0): 57204 runs - 18.35 us/run - 12320 kB/run - 640.57 GB/s
SOFT_MAX(type=f32,ne=[1024,1024,10,1],mask=0,sinks=0,m_prec=f32,nr23=[1,1],scale=1.000000,max_bias=0.000000,inplace=0): 9840 runs - 102.46 us/run - 81920 kB/run - 763.45 GB/s
SOFT_MAX(type=f32,ne=[77,1024,10,1],mask=0,sinks=0,m_prec=f32,nr23=[1,1],scale=1.000000,max_bias=0.000000,inplace=0): 98064 runs - 10.25 us/run - 6160 kB/run - 573.43 GB/s
SOFT_MAX(type=f32,ne=[256,256,20,1],mask=0,sinks=0,m_prec=f32,nr23=[1,1],scale=1.000000,max_bias=0.000000,inplace=0): 98310 runs - 10.25 us/run - 10240 kB/run - 953.20 GB/s
SOFT_MAX(type=f32,ne=[64,64,20,1],mask=0,sinks=0,m_prec=f32,nr23=[1,1],scale=1.000000,max_bias=0.000000,inplace=0): 172011 runs - 5.99 us/run - 640 kB/run - 101.84 GB/s
SOFT_MAX(type=f32,ne=[77,64,20,1],mask=0,sinks=0,m_prec=f32,nr23=[1,1],scale=1.000000,max_bias=0.000000,inplace=0): 172011 runs - 5.97 us/run - 770 kB/run - 123.02 GB/s
SOFT_MAX(type=f32,ne=[8192,1,1,1],mask=0,sinks=0,m_prec=f32,nr23=[1,1],scale=1.000000,max_bias=0.000000,inplace=0): 172011 runs - 6.00 us/run - 64 kB/run - 10.16 GB/s
SOFT_MAX(type=f32,ne=[8192,4,1,1],mask=0,sinks=0,m_prec=f32,nr23=[1,1],scale=1.000000,max_bias=0.000000,inplace=0): 163820 runs - 6.12 us/run - 256 kB/run - 39.91 GB/s
SOFT_MAX(type=f32,ne=[8192,16,1,1],mask=0,sinks=0,m_prec=f32,nr23=[1,1],scale=1.000000,max_bias=0.000000,inplace=0): 147438 runs - 6.88 us/run - 1024 kB/run - 141.92 GB/s
SOFT_MAX(type=f32,ne=[16384,1,1,1],mask=0,sinks=0,m_prec=f32,nr23=[1,1],scale=1.000000,max_bias=0.000000,inplace=0): 122865 runs - 8.20 us/run - 128 kB/run - 14.89 GB/s
SOFT_MAX(type=f32,ne=[16384,4,1,1],mask=0,sinks=0,m_prec=f32,nr23=[1,1],scale=1.000000,max_bias=0.000000,inplace=0): 114674 runs - 8.87 us/run - 512 kB/run - 55.06 GB/s
SOFT_MAX(type=f32,ne=[16384,16,1,1],mask=0,sinks=0,m_prec=f32,nr23=[1,1],scale=1.000000,max_bias=0.000000,inplace=0): 98292 runs - 10.24 us/run - 2048 kB/run - 190.82 GB/s
SOFT_MAX(type=f32,ne=[32768,1,1,1],mask=0,sinks=0,m_prec=f32,nr23=[1,1],scale=1.000000,max_bias=0.000000,inplace=0): 49146 runs - 21.37 us/run - 256 kB/run - 11.43 GB/s
SOFT_MAX(type=f32,ne=[32768,4,1,1],mask=0,sinks=0,m_prec=f32,nr23=[1,1],scale=1.000000,max_bias=0.000000,inplace=0): 49146 runs - 22.54 us/run - 1024 kB/run - 43.33 GB/s
SOFT_MAX(type=f32,ne=[32768,16,1,1],mask=0,sinks=0,m_prec=f32,nr23=[1,1],scale=1.000000,max_bias=0.000000,inplace=0): 49146 runs - 23.92 us/run - 4096 kB/run - 163.32 GB/s
SOFT_MAX(type=f32,ne=[65536,1,1,1],mask=0,sinks=0,m_prec=f32,nr23=[1,1],scale=1.000000,max_bias=0.000000,inplace=0): 32764 runs - 38.94 us/run - 512 kB/run - 12.54 GB/s
SOFT_MAX(type=f32,ne=[65536,4,1,1],mask=0,sinks=0,m_prec=f32,nr23=[1,1],scale=1.000000,max_bias=0.000000,inplace=0): 24573 runs - 41.94 us/run - 2048 kB/run - 46.57 GB/s
SOFT_MAX(type=f32,ne=[65536,16,1,1],mask=0,sinks=0,m_prec=f32,nr23=[1,1],scale=1.000000,max_bias=0.000000,inplace=0): 24582 runs - 43.09 us/run - 8192 kB/run - 181.32 GB/s
SOFT_MAX(type=f32,ne=[131072,1,1,1],mask=0,sinks=0,m_prec=f32,nr23=[1,1],scale=1.000000,max_bias=0.000000,inplace=0): 16382 runs - 74.56 us/run - 1024 kB/run - 13.10 GB/s
SOFT_MAX(type=f32,ne=[131072,4,1,1],mask=0,sinks=0,m_prec=f32,nr23=[1,1],scale=1.000000,max_bias=0.000000,inplace=0): 16382 runs - 79.85 us/run - 4096 kB/run - 48.92 GB/s
SOFT_MAX(type=f32,ne=[131072,16,1,1],mask=0,sinks=0,m_prec=f32,nr23=[1,1],scale=1.000000,max_bias=0.000000,inplace=0): 12294 runs - 82.41 us/run - 16384 kB/run - 189.64 GB/s
SOFT_MAX(type=f32,ne=[262144,1,1,1],mask=0,sinks=0,m_prec=f32,nr23=[1,1],scale=1.000000,max_bias=0.000000,inplace=0): 8191 runs - 145.16 us/run - 2048 kB/run - 13.46 GB/s
SOFT_MAX(type=f32,ne=[262144,4,1,1],mask=0,sinks=0,m_prec=f32,nr23=[1,1],scale=1.000000,max_bias=0.000000,inplace=0): 8194 runs - 155.46 us/run - 8192 kB/run - 50.26 GB/s
SOFT_MAX(type=f32,ne=[262144,16,1,1],mask=0,sinks=0,m_prec=f32,nr23=[1,1],scale=1.000000,max_bias=0.000000,inplace=0): 7175 runs - 160.70 us/run - 32768 kB/run - 194.56 GB/s
SOFT_MAX(type=f32,ne=[524288,1,1,1],mask=0,sinks=0,m_prec=f32,nr23=[1,1],scale=1.000000,max_bias=0.000000,inplace=0): 8191 runs - 285.81 us/run - 4096 kB/run - 13.67 GB/s
SOFT_MAX(type=f32,ne=[524288,4,1,1],mask=0,sinks=0,m_prec=f32,nr23=[1,1],scale=1.000000,max_bias=0.000000,inplace=0): 4098 runs - 306.91 us/run - 16384 kB/run - 50.92 GB/s
SOFT_MAX(type=f32,ne=[524288,16,1,1],mask=0,sinks=0,m_prec=f32,nr23=[1,1],scale=1.000000,max_bias=0.000000,inplace=0): 3591 runs - 317.06 us/run - 65536 kB/run - 197.32 GB/s
-- After
Backend 1/2: CUDA0
Device description: NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition
Device memory: 97250 MB (96691 MB free)
SOFT_MAX(type=f32,ne=[4096,4096,5,1],mask=0,sinks=0,m_prec=f32,nr23=[1,1],scale=1.000000,max_bias=0.000000,inplace=0): 2236 runs - 450.67 us/run - 655360 kB/run - 1400.15 GB/s
SOFT_MAX(type=f32,ne=[12888,256,5,1],mask=0,sinks=0,m_prec=f32,nr23=[1,1],scale=1.000000,max_bias=0.000000,inplace=0): 17748 runs - 56.97 us/run - 128880 kB/run - 2161.50 GB/s
SOFT_MAX(type=f32,ne=[77,4096,5,1],mask=0,sinks=0,m_prec=f32,nr23=[1,1],scale=1.000000,max_bias=0.000000,inplace=0): 57204 runs - 18.35 us/run - 12320 kB/run - 640.36 GB/s
SOFT_MAX(type=f32,ne=[1024,1024,10,1],mask=0,sinks=0,m_prec=f32,nr23=[1,1],scale=1.000000,max_bias=0.000000,inplace=0): 9840 runs - 102.46 us/run - 81920 kB/run - 763.42 GB/s
SOFT_MAX(type=f32,ne=[77,1024,10,1],mask=0,sinks=0,m_prec=f32,nr23=[1,1],scale=1.000000,max_bias=0.000000,inplace=0): 98064 runs - 10.25 us/run - 6160 kB/run - 573.43 GB/s
SOFT_MAX(type=f32,ne=[256,256,20,1],mask=0,sinks=0,m_prec=f32,nr23=[1,1],scale=1.000000,max_bias=0.000000,inplace=0): 98310 runs - 10.25 us/run - 10240 kB/run - 953.21 GB/s
SOFT_MAX(type=f32,ne=[64,64,20,1],mask=0,sinks=0,m_prec=f32,nr23=[1,1],scale=1.000000,max_bias=0.000000,inplace=0): 147438 runs - 7.00 us/run - 640 kB/run - 87.26 GB/s
SOFT_MAX(type=f32,ne=[77,64,20,1],mask=0,sinks=0,m_prec=f32,nr23=[1,1],scale=1.000000,max_bias=0.000000,inplace=0): 147438 runs - 6.99 us/run - 770 kB/run - 105.05 GB/s
SOFT_MAX(type=f32,ne=[8192,1,1,1],mask=0,sinks=0,m_prec=f32,nr23=[1,1],scale=1.000000,max_bias=0.000000,inplace=0): 172011 runs - 6.02 us/run - 64 kB/run - 10.13 GB/s
SOFT_MAX(type=f32,ne=[8192,4,1,1],mask=0,sinks=0,m_prec=f32,nr23=[1,1],scale=1.000000,max_bias=0.000000,inplace=0): 163820 runs - 6.12 us/run - 256 kB/run - 39.87 GB/s
SOFT_MAX(type=f32,ne=[8192,16,1,1],mask=0,sinks=0,m_prec=f32,nr23=[1,1],scale=1.000000,max_bias=0.000000,inplace=0): 147438 runs - 6.91 us/run - 1024 kB/run - 141.40 GB/s
SOFT_MAX(type=f32,ne=[16384,1,1,1],mask=0,sinks=0,m_prec=f32,nr23=[1,1],scale=1.000000,max_bias=0.000000,inplace=0): 122865 runs - 8.20 us/run - 128 kB/run - 14.89 GB/s
SOFT_MAX(type=f32,ne=[16384,4,1,1],mask=0,sinks=0,m_prec=f32,nr23=[1,1],scale=1.000000,max_bias=0.000000,inplace=0): 114674 runs - 8.79 us/run - 512 kB/run - 55.54 GB/s
SOFT_MAX(type=f32,ne=[16384,16,1,1],mask=0,sinks=0,m_prec=f32,nr23=[1,1],scale=1.000000,max_bias=0.000000,inplace=0): 98292 runs - 10.24 us/run - 2048 kB/run - 190.82 GB/s
SOFT_MAX(type=f32,ne=[32768,1,1,1],mask=0,sinks=0,m_prec=f32,nr23=[1,1],scale=1.000000,max_bias=0.000000,inplace=0): 131056 runs - 8.11 us/run - 256 kB/run - 30.12 GB/s
SOFT_MAX(type=f32,ne=[32768,4,1,1],mask=0,sinks=0,m_prec=f32,nr23=[1,1],scale=1.000000,max_bias=0.000000,inplace=0): 49146 runs - 22.54 us/run - 1024 kB/run - 43.33 GB/s
SOFT_MAX(type=f32,ne=[32768,16,1,1],mask=0,sinks=0,m_prec=f32,nr23=[1,1],scale=1.000000,max_bias=0.000000,inplace=0): 49146 runs - 23.32 us/run - 4096 kB/run - 167.50 GB/s
SOFT_MAX(type=f32,ne=[65536,1,1,1],mask=0,sinks=0,m_prec=f32,nr23=[1,1],scale=1.000000,max_bias=0.000000,inplace=0): 122865 runs - 8.19 us/run - 512 kB/run - 59.63 GB/s
SOFT_MAX(type=f32,ne=[65536,4,1,1],mask=0,sinks=0,m_prec=f32,nr23=[1,1],scale=1.000000,max_bias=0.000000,inplace=0): 40955 runs - 24.59 us/run - 2048 kB/run - 79.43 GB/s
SOFT_MAX(type=f32,ne=[65536,16,1,1],mask=0,sinks=0,m_prec=f32,nr23=[1,1],scale=1.000000,max_bias=0.000000,inplace=0): 24582 runs - 43.21 us/run - 8192 kB/run - 180.84 GB/s
SOFT_MAX(type=f32,ne=[131072,1,1,1],mask=0,sinks=0,m_prec=f32,nr23=[1,1],scale=1.000000,max_bias=0.000000,inplace=0): 122865 runs - 8.19 us/run - 1024 kB/run - 119.25 GB/s
SOFT_MAX(type=f32,ne=[131072,4,1,1],mask=0,sinks=0,m_prec=f32,nr23=[1,1],scale=1.000000,max_bias=0.000000,inplace=0): 40955 runs - 24.59 us/run - 4096 kB/run - 158.87 GB/s
SOFT_MAX(type=f32,ne=[131072,16,1,1],mask=0,sinks=0,m_prec=f32,nr23=[1,1],scale=1.000000,max_bias=0.000000,inplace=0): 12294 runs - 82.37 us/run - 16384 kB/run - 189.74 GB/s
SOFT_MAX(type=f32,ne=[262144,1,1,1],mask=0,sinks=0,m_prec=f32,nr23=[1,1],scale=1.000000,max_bias=0.000000,inplace=0): 122865 runs - 8.20 us/run - 2048 kB/run - 238.28 GB/s
SOFT_MAX(type=f32,ne=[262144,4,1,1],mask=0,sinks=0,m_prec=f32,nr23=[1,1],scale=1.000000,max_bias=0.000000,inplace=0): 36873 runs - 28.66 us/run - 8192 kB/run - 272.61 GB/s
SOFT_MAX(type=f32,ne=[262144,16,1,1],mask=0,sinks=0,m_prec=f32,nr23=[1,1],scale=1.000000,max_bias=0.000000,inplace=0): 9225 runs - 108.51 us/run - 32768 kB/run - 288.13 GB/s
SOFT_MAX(type=f32,ne=[524288,1,1,1],mask=0,sinks=0,m_prec=f32,nr23=[1,1],scale=1.000000,max_bias=0.000000,inplace=0): 98292 runs - 10.24 us/run - 4096 kB/run - 381.65 GB/s
SOFT_MAX(type=f32,ne=[524288,4,1,1],mask=0,sinks=0,m_prec=f32,nr23=[1,1],scale=1.000000,max_bias=0.000000,inplace=0): 32784 runs - 31.74 us/run - 16384 kB/run - 492.43 GB/s
SOFT_MAX(type=f32,ne=[524288,16,1,1],mask=0,sinks=0,m_prec=f32,nr23=[1,1],scale=1.000000,max_bias=0.000000,inplace=0): 8721 runs - 121.20 us/run - 65536 kB/run - 516.19 GB/s
```
* Fix compiler warnings by casting `const` away
* llama : require backend samplers to be of type llama_sampler_chain
* sampling : use host buffer type for inputs
* Try fixing HIP build errors by adding corresponding #defines
Will likely have to disable for MUSA as I didn't find any docs online
* Fix launch logic when supports_cooperative_launch=false
* Disable cooperative groups for musa
Didn't find any doc online, so I don't even know if they support this
* server : reconnect the backend_sampling setting in the WebUI
* graph : make the compute graph constant with respect to active samplers
* batch : fix sequence id ownage
* graph : respect sampler order for graph reuse
* HIP/MUSA: fix build for backend sampling
* sampling : optimize logit_bias sampler
* cont : fix build
* sampling : generic ggml op support detection
* sampling : fix greedy
* tests : run backend sampler tests always on the CPU
* Apply suggestions from code review
Co-authored-by: Johannes Gäßler <johannesg@5d6.de>
* webui : fix lint
* Fix data-race in `soft_max_f32_parallelize_cols_single_row`
By using `tmp_vals` to store both max values and exponential
accumulator there was a potential data-race, where the exponential accumulator
for a given CTA may have written to `tmp_vals` before all others CTAs have
read the max value from it.
To avoid a third g.sync(), an additional temporary data-storage was
added. Given that there are syncs in place after writing to gmem, it is
guaranteed that the previous values for sums/max were read by all CTAs now.
* Apply automated code-formating to softmax.cu
* llama : clarify backend_accept/backend_set_input comments [no ci]
* llama : fix typo in comment [no ci]
* tests : use smart pointers for backend samplers
* tests : use smart pointers for model and context
* tests : remove vocab member from test_model_context
Also includes some minor cleanups related to nullptr checks.
* tests : extract batch info update to separate method
* tests : fix batch token position tracking in test_backend_sampler.cpp
* tests : add --device option support to backend sampler tests
This commit adds support for specifying a device to run the test on.
* common : disable backend sampling when grammar is involved
* Fix different RNG-states between backend-sampling and llama-sampling
By default, we perform a warm-up step where the ggml_cgraph is computed
once. For backend-sampling, this graph contains the sampler, and thus
the RNG state of the backend's dist sampler is advanced once.
Solution to this is to reset the samplers after the warmup has finished
* Make backend dist sampler use same rnd's as dist sampler
We sample in double precision and cast to float to match rnd numbers of
llama_dampler_dist which uses double precision (sampling from
std::uniform_real_distribution<double> and
std::uniform_real_distribution<float> with same rng will produce
different sequences).
* Update CCCL version to v3.2.0-rc2
* Build with CCCL 3.2 for CUDA backends
Gives best perf for backend-sampling on CUDA. Flag can be removed once
CCCL 3.2 is bundled within CTK and that CTK version is used in llama.cpp
* tests : revert server test changes (no longer needed)
* ggml : include cub/cub.cuh instead of block_scan.cuh
This commit updates the include directive in cumsum.cu to use
cub/cub.cuh instead of cub/block/block_scan.cuh.
The motivation of this change is that without it compilation fails
with the following error:
```console
/llama.cpp/ggml/src/ggml-cuda/cumsum.cu(196): error: name followed by "::" must be a class or namespace name
cub::DeviceScan::InclusiveSum(nullptr,
^
/llama.cpp/ggml/src/ggml-cuda/cumsum.cu(207): error: name followed by "::" must be a class or namespace name
cub::DeviceScan::InclusiveSum((void *) tmp_alloc.get(), tmp_size, src, dst, ne, stream);
^
2 errors detected in the compilation of "/llama.cpp/ggml/src/ggml-cuda/cumsum.cu".
gmake[2]: *** [ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/build.make:317: ggml/src/ggml-cuda/CMakeFiles/ggml-cuda.dir/cumsum.cu.o] Error 2
```
Commit
|
||
|
Georgi Gerganov
|
c560316440 |
graph : reuse SSM graphs (#16490)
* graph : reuse hybrid graphs * graph : reuse recurrent graphs * graph : fix reuse check for recurrent inputs * memory : move the recurrent state into the memory context * Revert "memory : move the recurrent state into the memory context" This reverts commit 00f115fe810815d4a22a6dee0acc346131e970e1. * cont : fix build |
||
|
Xuan-Son Nguyen
|
0759b09c90 | graph: add f_attn_temp_offset (#18025) | ||
|
Georgi Gerganov
|
e38b7c6e9e |
graph : support cacheless embeddings with FA and iSWA (#16528)
* graph : support cacheless embeddings with FA and iSWA * cont : deduplicate mask creation * cont : fix name |
||
|
|
e08db42595 |
model: EmbeddingGemma Adding Support for SentenceTransformers Dense Modules (#16367)
* model: EmbeddingGemma sentence-transformers dense linear projections support * model: add support for EmbeddingGemma SentenceTransformers dense linear projections Adding support for the Dense modules used in EmbeddingGemma models. EmbeddingGemma is a SentenceTransformers model with additional modules beyond the base Transformer backbone. See: https://developers.googleblog.com/en/gemma-explained-embeddinggemma-architecture-and-recipe/ * model: add support for EmbeddingGemma SentenceTransformers dense linear projections - converting model with dense-layers is optional - introduced dense config params * Update convert_hf_to_gguf.py Co-authored-by: Daniel Bevenius <daniel.bevenius@gmail.com> * fixed formatting issues * Update src/llama-graph.cpp Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * - removed pooling_type_opt, always allow overriding pooling_type - asserts checking dense features dims * fix python lint * fix ubuntu gcc build warning * - fixed thread-safety test - moved asserts to load_hparams * - tidying up code - simplifying graph-context expecting both dense weights * minor : add TODO --------- Co-authored-by: Daniel Bevenius <daniel.bevenius@gmail.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> |
||
|
Douglas Hanley
|
b5bd037832 | llama : add support for qwen3 reranker (#15824) | ||
|
Daniel Bevenius
|
fb15d649ed |
llama : add support for EmbeddingGemma 300m (#15798)
This commit add support for the EmbeddingGemma 300m. This model supports sliding window attention (SWA) and a new swq_type is introduced to support symmetric SWA masking. This commit also extracts the code from the function llama_is_masked_swa in llama-impl.h, so that the logic can be shared by both llm_graph_input_attn_no_cache::set_input and llama_kv_cache::set_input_kq_mask. With this commit the EmbeddingGemma 300m model can be converted to to GGUF and used with llama.cpp. Once the model has been uploaded to HuggingFace it can be used like this: ```console ./build/bin/llama-cli -hf ggml-org/embeddinggemma-300m-GGUF:Q8_0 ``` |
||
|
Johannes Gäßler
|
e81b8e4b7f |
llama: use FA + max. GPU layers by default (#15434)
* llama: use max. GPU layers by default, auto -fa * ggml-backend: abort instead of segfault |
||
|
Georgi Gerganov
|
3f196be84b |
graph : remove build_attn_with_sinks overload (#15469)
ggml-ci |
||
|
Georgi Gerganov
|
715a6db02c |
kv-cache : drop the "unified" prefix (#15467)
* kv-cache : drop the "unified" prefix ggml-ci * cont : fix comment [no ci] |
||
|
|
fd1234cb46 |
llama : add gpt-oss (#15091)
* oai moe * compat with new checkpoint * add attn sink impl * add rope scaling yarn * logits match with latest transformers code * wip chat template * rm trailing space * use ggml_scale_bias * rm redundant is_swa_all * convert interleaved gate_up * graph : fix activation function to match reference (#7) * vocab : handle o200k_harmony special tokens * ggml : add attention sinks support (#1) * llama : add attn sinks * ggml : add attn sinks * cuda : add attn sinks * vulkan : add support for sinks in softmax remove unnecessary return * ggml : add fused swiglu_oai op (#11) * ggml : add fused swiglu_oai op * Update ggml/src/ggml-cpu/ops.cpp Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * update CUDA impl * cont : metal impl * add vulkan impl * test-backend-ops : more test cases, clean up * llama : remove unfused impl * remove extra lines --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> --------- Co-authored-by: slaren <slarengh@gmail.com> * repack mxfp4 upon conversion * clean up a bit * enable thinking * add quick hack to render only some special tokens * fix bf16 conversion * remove vocab hack * webui ok * support chat parsing for gpt-oss * fix webui * direct mapping mxfp4, FINALLY * force using mxfp4 * properly use lazy tensor * ggml : add mxfp4 ggml : use e8m0 conversion instead of powf Co-authored-by: Diego Devesa <slarengh@gmail.com> change kvalues_mxfp4 table to match e2m1 (#6) metal : remove quantization for now (not used) cuda : fix disabled CUDA graphs due to ffn moe bias vulkan : add support for mxfp4 cont : add cm2 dequant * ggml : add ggml_add_id (#13) * ggml : add ggml_add_id * add cuda impl * llama : add weight support check for add_id * perf opt * add vulkan impl * rename cuda files * add metal impl * allow in-place ggml_add_id * llama : keep biases on CPU with --cpu-moe * llama : fix compile error ggml-ci * cuda : add fallback for __nv_cvt_e8m0_to_bf16raw ggml-ci * cleanup ggml-ci * sycl : fix supports_op for MXFP4 ggml-ci * fix Unknown reasoning format * ggml-cpu : fix AVX build ggml-ci * fix hip build ggml-ci * cuda : add mxfp4 dequantization support for cuBLAS ggml-ci * ggml-cpu : fix mxfp4 fallback definitions for some architectures ggml-ci * cuda : fix version required for __nv_cvt_e8m0_to_bf16raw --------- Co-authored-by: Xuan Son Nguyen <son@huggingface.co> Co-authored-by: slaren <slarengh@gmail.com> |
||
|
Georgi Gerganov
|
ba42794c9e |
graph : fix equal_seq() check (#14986)
ggml-ci |
||
|
Dongliang Wei
|
c1dacaa99b | llama : merge build_moe_ffn_from_probs function into build_moe_ffn (#14968) | ||
|
compilade
|
66625a59a5 |
graph : reduce splits for recurrent and hybrid models (#14825)
* graph : avoid creating redundant s_copy views * graph : comment the s_copy views |
||
|
Georgi Gerganov
|
1e15bfd42c |
graph : fix stack-use-after-return (#14960)
ggml-ci |
||
|
|
6c6e397aff |
model : add support for SmallThinker series (#14898)
* support smallthinker * support 20b softmax, 4b no sliding window * new build_moe_ffn_from_probs, and can run 4b * fix 4b rope bug * fix python type check * remove is_moe judge * remove set_dense_start_swa_pattern function and modify set_swa_pattern function * trim trailing whitespace * remove get_vocab_base of SmallThinkerModel in convert_hf_to_gguf.py Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com> * better whitespace Apply suggestions from code review Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com> * use GGML_ASSERT for expert count validation Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com> * Improve null pointer check for probs Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com> * use template parameter for SWA attention logic * better whitespace Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * move the creation of inp_out_ids before the layer loop * remove redundant judge for probs --------- Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> |
||
|
Georgi Gerganov
|
8f974bc1e9 |
graph : refactor context to not pass gf explicitly (#14629)
ggml-ci |
||
|
Georgi Gerganov
|
01612b7409 |
llama : reuse compute graphs (#14482)
* llama : reuse compute graphs ggml-ci * llama-bench : add graph reuse parameter ggml-ci * cont : remove the parameter and the sched resets ggml-ci * graph : rename update() to can_reuse() ggml-ci * params : remove is_same() ggml-ci * graph : set res->params in llm_graph_context constructor ggml-ci * graph : avoid set_max_nodes in llm_graph_result ggml-ci * kv-cache : reuse llama_context's graph result instance ggml-ci * context : reset the previous graph result upon memory updates ggml-ci * batch : llama_ubatch now carries its data instead of pointing to balloc ggml-ci * merge : fix build ggml-ci * graph : fix can_reuse() checks when flash-attention is disabled * graph : move llm_graph_result impl in source file + debug env ggml-ci |
||
|
Georgi GerganovandJohannes Gäßler
|
225e7a1438 |
llama : add high-throughput mode (#14363)
* kv-cache : prepare K/V buffers for separation ggml-ci * batched-bench : fix oob write ggml-ci * llama : add "virtual sequences" ggml-ci * llama : use "stream" vs "virtual sequence" ggml-ci * graph : fix stream splitting when KV cache is not used ggml-ci * kv-cache : add multi-stream save/load support ggml-ci * llama : add "--attn-streams" flag ggml-ci * kv-cache : fix handling when find_slot fails ggml-ci * kv-cache : restore find_slot impl ggml-ci * kv-cache : add comments * kv-cache : add bounds checks for sequence id ggml-ci * cont : add n_seq_max to batch allocr ggml-ci * kv-cache : perform stream copies lazily after llama_synchronize ggml-ci * kv-cache : avoid throwing exceptions across the C boundary ggml-ci * CUDA: 4D FlashAttention support (#14628) * CUDA: 4D FlashAttention support * CUDA: fix WMMA FA kernel * llama : rename attn_streams -> kv_unified ggml-ci * common : rename kv_split -> kv_unified ggml-ci --------- Co-authored-by: Johannes Gäßler <johannesg@5d6.de> |
||
|
Xuan-Son Nguyen
|
cb9178f885 | llama : remove llm_graph_input_one (#14603) | ||
|
compiladeandSigbjørn Skjæret
|
4a5686da22 |
llama : support Jamba hybrid Transformer-Mamba models (#7531)
* wip: llama : separate recurrent states from the KV cache This will be necessary to support Jamba (and other recurrent models mixed with Attention). Doesn't compile yet, and finding a slot isn't yet done correctly for recurrent states. * llama : use std::find for seq_nodes in llama_rs_cache * llama : state checkpoints for recurrent models * llama : correctly handle more edge cases for the rs cache * llama : rename many llama_kv_cache_* functions * llama : remove useless return value for some llama_cache_* functions * llama : rethink recurrent state cell counts * llama : begin work on support for variable GQA This will also be useful for Jamba if we consider the Mamba layers to have 0 KV heads. * llama : gracefully fail when not finding hybrid slot * llama : support Jamba * llama : fix BERT inference without KV cache * convert-hf : check for unprocessed Jamba experts * convert-hf : support Mini-Jamba conversion * llama : fix Jamba quantization sanity checks * llama : sequence-length-aware batch splitting * llama : use equal-sequence-length sub-batches for recurrent models * ggml : simplify SSM-related operators * llama : make recurrent state slot allocation contiguous * llama : adapt internal uses of batches to llama_ubatch * llama : fix batch split output count for embeddings * llama : minimize swaps when reordering logits This reduces overhead when running hellaswag on thousands of sequences with very small 100k params Mamba models. * llama : fix edge case finding batch seq_id of split recurrent cell This otherwise was a problem when running the HellaSwag benchmark with small batch sizes, making it crash. * llama : avoid copies for simple batch splits * ggml : make ggml_ssm_scan not modify its source tensors * llama : fix shared recurrent tail cell count for small ubatch sizes Otherwise it was impossible to run the 'parallel' example with '-ub 1' with a Mamba or Jamba model. * llama : fix .base() compilation error on Windows * llama : allow doing the equivalent of SSM_CONV with SUM_ROWS and MUL * ggml : allow GGML_OP_CONCAT to work on non-contiguous tensors The implementation already supported it, and this makes Mamba's conv step slightly faster. * mamba : fix non-contiguous usage of ggml_silu * llama : session saving and reloading for hybrid models * convert_hf : fix Jamba conversion * llama : fix mixed signedness comparison * llama : use unused n_embd_k_gqa in k_shift This also slightly reduces the diff from the master branch * llama : begin renaming llama_past back to llama_kv_cache * llama : remove implicit recurrent state rollbacks * llama : partially apply clang-format style * convert : fix jamba conv1d shape squeezing * graph : add back hybrid memory graph input But this time it contains the sub-cache graph inputs. This *should* make it easier to handle updating the inputs when caching the graph (eventually). * model : add Jamba to Mamba-specific hparams printing * jamba : remove redundant nullptr initializations * model : remove unnecessary prefix for tensor loading constants Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com> * model : use ggml_swiglu_split for Mamba Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com> * model : make falcon-h1 use shared mamba2 layer builder * memory : avoid referring to KV in recurrent cache logs * gguf-py : avoid adding duplicate tensor mappings for Jamba Some of the tensor names are common with Llama4 --------- Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com> |
||
|
Georgi Gerganov
|
7b50f7c025 |
graph : prepare for 4D mask (#14515)
ggml-ci |
||
|
Georgi Gerganov
|
a70c8a0c4b |
kv-cache : use ggml_set_rows (#14285)
* kv-cache : use ggml_set_rows ggml-ci * graph : separate k and v indices ggml-ci * cont : remove redundant ifs ggml-ci * kv-cache : improve find_slot impl * kv-cache : bounds-check when accessing slot_info indices * kv-cache : add comments ggml-ci * ggml : add TODOs for adding GGML_OP_SET_ROWS support in the backends ggml-ci |
||
|
compilade
|
5d46babdc2 |
llama : initial Mamba-2 support (#9126)
* llama : initial Mamba-2 support * ggml : SIMD ggml_ssm_scan for Mamba-2 * ggml : improve ggml_mul speed when masking recurrent states * llama : support running Mamba-Codestral-7B-v0.1 * llama : fix Mamba-2 conv state saving * ggml : make the ggml_mul fast broadcast path more consistently formatted * llama : remove unused variable * llama : add missing break * convert_hf : prefer SentencePiece tokenizer for Mamba-2 when present The tokenzier.json of Mamba-Codestral-7B-v0.1 otherwise requires workarounds to work correctly. * llama : avoid redundant state copy for Mamba 1 and 2 * metal : attempt to adapt SSM_SCAN for Mamba-2 * metal : fix SSM_SCAN pipeline scope * metal : use log and exp instead of log1pf and expf in SSM_SCAN * metal : remove unused arguments for SSM_SCAN The max index is 31, so trimming the arguments is necessary. * metal : add back n_seqs to SSM_SCAN args Whoops, this is needed for the offset in the concatenated output. * metal : fix SSM_SCAN state head offset * metal : fix wrong number of tokens per sequence in SSM_SCAN * ggml : remove unused fast broadcast path in GGML_MUL This was initially added because states were masked with ggml_mul, but this is no longer done and so this "optimisation" is no longer necessary, or at least not worth the additional code complexity. * ggml : avoid multiply by D in GGML_OP_SSM_SCAN This makes the weight buft detection in src/llama.cpp simpler. * convert : transpose Mamba-2 A, D and reshape SSM_NORM This breaks existing conversions of Mamba-2 models to avoid some reshapes. Not sure if it's a good idea, but it makes the graph slightly cleaner. * llama : more appropriate SSM_SCAN and SSM_CONV buft support checks * convert : fix flake8 lint * metal : fix confusion between ; and , * metal : add missing args for nb references in ssm_scan_f32_group * metal : single-user mamba2 inference works * kv-cache : remove const_cast when setting inputs for s_copy And also fix multi-user inference for recurrent models by using cell_id instead of i as the kv cell index when populating s_copy. * convert : avoid AutoConfig for Mamba and Mamba2 hparams * kv-cache : allow context shift for recurrent models * graph : fix recurrent state copies when avoiding copies Works, but using lambda functions might not be that clean. * ggml : fix mamba2 ssm scan when compiled with SVE * ggml-cpu : reorder SVE FMA for consistency with other SIMD arches * cuda : implement ssm scan for Mamba2 There is still room for improvement, but it works! * cuda : adapt Mamba1 ssm scan to shape changes from Mamba2 * mamba : fix mismatched new and delete size for llm_build_mamba Subclasses of llm_graph_context cannot have extra fields, because the called destructor is not the one from the subclass. This otherwise would cause problems when runnning Mamba-(1|2) inference when compiled -DGGML_SANITIZE_ADDRESS=ON * cuda : graceful fallback for Mamba-1 models with weird embd size |