* Copy reduce result to other GPUs if necessary
* Avoid ggml_get_rows for TG
* For the output ops use the result of the split that ran on the main GPU
* More models
* A hopefully more efficient adaptive_p sampling
* Once at it, lets fix the formatting too
* More formatting
* Hopefully better
* This should be better
* Correctly accumulate adaptive_p sampling time
* AVX2
* Attempt to fix the many GPU issue in split mode graph
* WIP: this seems more stable
Still hanging after a while if I try to use all 7 GPUs
* Reenable OpenMP in scheduler async
Seems solid up to 4 GPUs. It did hang with --max-gpu 6.
* printf cleanup
* WIP - not working
* WIP - not working
* WIP - GPT-OSS working
However, extremely stupid. The only way I could correctly repack the
up/gate experts is to copy up and gate into host buffers, repack
into another host buffer, copy back into the ffn_up_gate_exps tensor.
This is going to be very slow for giant 500 GB models.
My attempts to do this via a compute graph on the backend holding
the tensors was unsuccessful.
For GPT-OSS-20B I see ~6-7% better PP when using the original
ik_llama.cpp fused_up_gate CUDA implementation, and ~10% when
using the small batch size implementation.
Other models are not working yet on CUDA as I need to fix the
fused mul-unary implementation.
* WIP
* WIP - Qwen3-MoE (and hopefully all others) working
But when I say here and in the previous commit "working",
I mean PP is working. TG is still broken.
* WIP: TG seems to be working
* Minor
* Add command line option to merge experts up/gate
* Add merge up/gate command line parameter to llama-bench
* Turn off merge_up_gate_exps if split mode graph
It is not yet implemented
* When no bias, allow merging up/gate with tensor overrides
* Arghh, we need to increase the context size again
* Cleanup
* server: improve speed of speculative decoding
change logs
rpc: add recompute
spec dec fix
* Fix n_batch_size not set to context size for draft model
---------
Co-authored-by: firecoperana <firecoperana>
* Mimo-2 support
* Fix bug for head sizes not being the same
It still does not solve the Mimo-2 quantized cache issue.
* Fix quantized cache
* Minor
---------
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
* WIP: absorb adding input into std_attn and std_ffn
* WIP: NCCL infra
* WIP: add reduce and fake_cpy ops
* WIP
* WIP: graph appears to work, layer is broken
* WIP: Qwen3-MoE works with graph, layer still broken
* WIP: GLM-4.5 graph works
* WIP: fix sm layer (dense)
* WIP: fix sm layer (MoE)
* WIP: fast PP with bespoke 4-GPU NCCL
I guess, I'm not using NCCL the right way as PP is very
low with a single communicator group for 3 or more GPUs.
But if I create 4 communicator groups for pairs of GPUs
(0,1, 2,3, 0,2, 1,3) and use that, PP is fast: I'm hitting
1500 t/s for L3-70B on the 4x3090 system, which is
~20% better than the previous sm graph without NCCL.
But that cannot be the solution (I cannot be creating pairwise
communicators and associated logic for every possible number of GPUs).
* WIP: Cohere2
* Explicitely set device
* Bespoke 3-GPU case
* WIP
* Do not repeat get_rows multiple times
* Fix 3 GPUs
* OK, let's leave it in
* Simple async
* This sync seems enough
* Only do async for 4 or more backends
With 2 GPUs (so, 3 backends) not using async is slightly faster
* Scheduler changes
* Use OpenMP if available
Surprisingly (at least to me), this is quite a bit faster than
std::thread and std::barrier. GLM-4.5-AIR with 4 GPUs is now
at 105 t/s at zero context!
* Do not use OpenMP if there are tensor overrides
* Set omp max active levels
* Be more careful with having set the device before using a stream
* Command line option to turn on async. Set to false by defualt for now
---------
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
* Adding fused_norm - same idea as fused_rms_norm
* Avoid computing the attention reduce op for cohere2
---------
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
* WIP: absorb adding input into std_attn and std_ffn
* WIP: NCCL infra
* WIP: add reduce and fake_cpy ops
* WIP
* WIP: graph appears to work, layer is broken
* WIP: Qwen3-MoE works with graph, layer still broken
* WIP: GLM-4.5 graph works
* WIP: fix sm layer (dense)
* WIP: fix sm layer (MoE)
* WIP: fast PP with bespoke 4-GPU NCCL
I guess, I'm not using NCCL the right way as PP is very
low with a single communicator group for 3 or more GPUs.
But if I create 4 communicator groups for pairs of GPUs
(0,1, 2,3, 0,2, 1,3) and use that, PP is fast: I'm hitting
1500 t/s for L3-70B on the 4x3090 system, which is
~20% better than the previous sm graph without NCCL.
But that cannot be the solution (I cannot be creating pairwise
communicators and associated logic for every possible number of GPUs).
* WIP: Cohere2
* Explicitely set device
* Bespoke 3-GPU case
* WIP
* Do not repeat get_rows multiple times
* Fix 3 GPUs
* OK, let's leave it in
* Implement the reduce op without NCCL available
* Be able to build without NCCL
cmake -DGGML_NCCL=OFF disables it
* Make --max-gpu work again
* Slightly better for 4 GPUs without NCCL
* Cleanup
---------
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
* This should do the trick for PP
* Command line option to set max. extra VRAM that the scheduler can use
* Fix bug and cleanup
* Looks like with this change it is working with tensor overrides
* Nah, it is not working
* OK, this seems to be working
* Disable split scheduling with tensor overrides
---------
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
* This works and TG is descent, but PP is low
* Better
* Apply f_logit_scale before mul mat with output tensor
* This is better for PP: 600 t/s -> 700 t/s
* To not lose this again
* WIP
* Equal split
* WIP
---------
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
* Rearrange graph nodes
So that we can do graph portions that are the same on 2 or more
GPUs at the same time.
* Separate graph compute implementation for split mode graph
* This is better
---------
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
* Remove most of split mode row
* WIP
* WIP: also allocate the KV cache using tensor split
* WIP: it runs with wrong result
But it also looks like the backend scheduler is not going to help:
* It copies mask and input positions to GPU 0
* => RoPE ops must run on GPU 0
* => To proceed attn evaluation, GPU 1 must wait for GPU 0 to finish its
entire attn calculation
* Same with FFN. The rms_norm gets scheduled on GPU 0. Hence, GPU 1 must
wait for GPU 0 to finish its entore FFN calculation before it can
start (as it needs to copy the result of rms_norm from GPU 0)
* => Seems useless without writing a bespoke TP scheduling
* WIP
* This works, but it is slow
* This is slightly better
the graph is still not being computed in parallel.
Why? Because the scheduler creates graph splits where the
result of the computation on one GPU becomes an input for the
other split. Hence, to trigger the computation on the second GPU
one needs to wait for the computation on the first GPU to finish,
even thiough the two can be done in parallel up to the sunchronization
point. So, all that is left to do is to trick the scheduler to create
to splits that can be done in parallel, and then have a graph split
where the results get combined.
* Playing games with the scheduler
This change tricks it into doing the right thing^TM.
Still quite a bit slower than split mode layer for the 8B LlaMA model.
But for the 70B LlaMA it now beats split mode layer for TG:
28 t/s vs 24.4 t/s. PP is 627 t/s vs 744 t/s.
In comparison, split mode "row" in mainline gets
484 t/s PP and 19.3 t/s TG.
* Fix attn split
Granularity for Wq, Wo is not just head size, but
head size * gqa_ratio.
Else the Wk, Wv tensors end up not being a multiple of the
head size when we divide the split determined by Wo with
the gqa_ratio.
* Show memory used per device
* Make it work with partial offload
but no tensor overrides yet, just ngl < num_layers.
* Allow for f16 source in fused_rms_norm
* This results in faster PP.
Now PP is faster than split mode layer for L3-70B.
* Rename split mode "row" to split mode "graph"
* Leave FFN partial results as f16
* WIP GLM4.5 - runs with wrong results
* WIP GLM4.5 - this works
PP is already better than split mode layer, but TG for zero context
is kind of low - 60 vs 92 t/s. TG becomes better than split mode layer
at around 20k tokens. PP at 26k tokens is 1.55X of sm layer.
* Work around compiler bug
It issues a warning that there is an extra semicolon outside of a function,
but there isn't. If I remove the anonymous namespace and turn the
functions inside into static, the warning disapears, so clearly
a compiler bug.
* Make graph reuse work with split mode graph
* Remove more split mode row remnants
* WIP tensor overrides
Runs with wrong results, don't see where the issue could be.
* This works but is slow
Still does not work for row-interleaved quants
* Slightly better
* Slightly better
* Row-interleaved quants work
* Better
* Minor
* Guarad against using split mode "graph" for unsupported models
* Guards against using merge_qkv with split mode "graph"
* WIP split mode attn
Works for LlaMA models, but not for GLM-4.5.
Doesn't seem to improve performance, so I guess no point in trying to
fix it.
* Split mode graph for qwen3moe
* Try to better distribute the splits
---------
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
* RPC support multiple devices
* rpc : update documentation (#16441)
Update the README file to match the newly added functionality of
exposing multiple devices from a single server.
Co-authored-by: Diego Devesa <slarengh@gmail.com>
# Conflicts:
# examples/rpc/README.md
* Remove memory settings
* rpc : cache and reuse compute graphs (#15405)
Store the last computed graph and reuse it when possible.
Also do not return response from GRAPH_COMPUTE and assume it always
completes successfully. If this this is not the case, the server closes
the connection. This saves us a network round trip to the server.
* Add -cpu to include cpu backend
---------
Co-authored-by: firecoperana <firecoperana>
Co-authored-by: Radoslav Gerganov <rgerganov@gmail.com>
* model : Granite docling + Idefics3 preprocessing (SmolVLM) (#16206)
* feat: Add granite-docling conversion using trillion pretokenizer
Branch: gabe-l-hart/GraniteDocling
Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
* feat: Add granite-docling vocab pre enum
Branch: gabe-l-hart/GraniteDocling
Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
* fix: Use granite-docling pre
Branch: gabe-l-hart/GraniteDocling
Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
* feat: Add clip_is_idefics3
Branch: gabe-l-hart/GraniteDocling
Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
* feat: Allow multi-token boundary sequences for image templating
Branch: gabe-l-hart/GraniteDocling
Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
* feat: Add tiling support for idefices3 in clip.cpp
This should likely be moved into llava_uhd::get_slice_instructions, but for
now this avoids disrupting the logic there.
Branch: gabe-l-hart/GraniteDocling
Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
* feat: Partial support for full templating for idefics3 in mtmd
There are still errors encoding some of the image chunks, but the token
sequence now matches transformers _almost_ perfectly, except for the double
newline before the global image which shows up as two consecutive newline
tokens instead of a single double-newline token. I think this is happening
because the blocks are tokenized separately then concatenated.
Branch: gabe-l-hart/GraniteDocling
Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
* feat: Fully working image preprocessing for idefics3 w/ resize and slicing
Branch: gabe-l-hart/GraniteDocling
Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
* feat: Parse the preprocessor config's longest side and add it to the mmproj hparams
Branch: GraniteDocling
Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
* fix: Use the longest side instead of size * scale_factor
For Granite Docling, these come out to the same value, but that was just a
conicidence.
Branch: GraniteDocling
Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
* fix: Allow batch encoding and remove clip_is_idefics3
Branch: GraniteDocling
Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
* refactor: Remove unnecessary conditionals for empty token vectors
Branch: GraniteDocling
Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
* refactor: Use image_manipulation util
Branch: GraniteDocling
Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
* add test model
---------
Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
Co-authored-by: Xuan Son Nguyen <son@huggingface.co>
# Conflicts:
# convert_hf_to_gguf.py
# convert_hf_to_gguf_update.py
# gguf-py/gguf/constants.py
# gguf-py/gguf/gguf_writer.py
# src/llama-vocab.cpp
# src/llama-vocab.h
* mtmd : support home-cooked Mistral Small Omni (#14928)
* model : add LightOnOCR-1B model (#16764)
* model : add LightOnOCR-1B model
* add test
# Conflicts:
# convert_hf_to_gguf.py
# gguf-py/gguf/constants.py
* mtmd : fix idefics3 preprocessing (#16806)
* mtmd : fix idefics3 preprocessing
* disable granite test
* fix test for granite
* model: Add support for CogVLM model (#15002)
* Added GGUF mappings for CogVLM model
* Add tensor mapping for CogVLM visual encoder
* Add CogVLM to conversion script, no vision part yet
* Added CogVLM vision model to conversion script
* Add graph for CogVLM CLIP model
* Add graph for CogVLM
* Fixes for CogVLM. Now compiles.
* Model now runs
* Fixes for cogvlm graph
* Account for graph context change after rebase
* Changes for whitespace
* Changes in convert script according to comments
* Switch CogVLM LLM graph to merged QKV tensor
* Use rope_type variable instead of direct definition
* Change CogVLM CLIP encoder to use SWIGLU
* Switch CogVLM CLIP to use merged QKV
* Apply rebase edits and remove ggml_cont call that is now unnecessary
* clean up
---------
Co-authored-by: Xuan Son Nguyen <son@huggingface.co>
# Conflicts:
# convert_hf_to_gguf.py
# examples/mtmd/clip.cpp
# gguf-py/gguf/constants.py
# gguf-py/gguf/tensor_mapping.py
# src/llama-arch.cpp
# src/llama-arch.h
# src/llama-model.cpp
# src/llama-model.h
* mtmd: refactor preprocessing + support max/min pixels (#16878)
* mtmd: refactor preprocessing + support max/min pixels
* fix mlp type
* implement mix/max pixels
* improve hparams
* better image preproc for qwen
* fix
* fix out of bound composite
* fix (2)
* fix token calculation
* get_merge_kernel_size()
* fix llama4 and lfm2
* gonna fix them all
* use simple resize for qwen
* qwen: increase min tokens
* no resize if dst size == src size
* restore to initial min/max tokens value for qwen
# Conflicts:
# examples/mtmd/clip.cpp
* clip : use FA (#16837)
* clip : use FA
* cont : add warning about unsupported ops
* implement "auto" mode for clip flash attn
* clip : print more detailed op support info during warmup
* cont : remove obsolete comment [no ci]
* improve debugging message
* trailing space
* metal : remove stray return
---------
Co-authored-by: Xuan Son Nguyen <son@huggingface.co>
* model: add Janus Pro for image understanding (#16906)
* Add support for Janus Pro
* Update gguf-py/gguf/tensor_mapping.py
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>
* Update gguf-py/gguf/tensor_mapping.py
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>
* Address reviewer suggestions
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>
* Add JANUS_PRO constant
* Update clip model handling
Co-authored-by: Xuan-Son Nguyen <son@huggingface.co>
* Update tools/mtmd/clip.cpp
Co-authored-by: Xuan-Son Nguyen <thichthat@gmail.com>
* Refactor JANUS_PRO handling in clip.cpp
Co-authored-by: Xuan-Son Nguyen <son@huggingface.co>
* Update tools/mtmd/clip.cpp
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>
* em whitespace
---------
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>
Co-authored-by: Xuan-Son Nguyen <son@huggingface.co>
Co-authored-by: Xuan-Son Nguyen <thichthat@gmail.com>
# Conflicts:
# convert_hf_to_gguf.py
# gguf-py/gguf/constants.py
# gguf-py/gguf/tensor_mapping.py
* mtmd: pad mask for qwen2.5vl (#16954)
* mtmd: pad mask for qwen2.5vl
* improve
* mtmd: add --image-min/max-tokens (#16921)
* mtmd: improve struct initialization (#16981)
* mtmd: allow QwenVL to process larger image by default (#17020)
* Disable flash attention
* mtmd : fix embedding size for image input (#17123)
* mtmd: fix patch_size initialized to random value in audio models (#17128)
* mtmd: fix patch_size initialized to random value in audio models
* add default hparams
* add llama_model_n_embd_inp
* Fix load qwen3 vl
Change batch size
* Add description
* Fix cli build error
---------
Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
Co-authored-by: Gabe Goodhart <ghart@us.ibm.com>
Co-authored-by: Xuan Son Nguyen <son@huggingface.co>
Co-authored-by: Tianyue-Zhao <zhaotianyue@outlook.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
Co-authored-by: Zhiyong Wang <85110830+ravenouse@users.noreply.github.com>
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>
Co-authored-by: Xuan-Son Nguyen <thichthat@gmail.com>
Co-authored-by: firecoperana <firecoperana>
Shortfixes the bug : ggml\src\ggml-cuda\cpy.cu:614: ggml_cuda_cpy_fn: unsupported type combination (q6_0 to f16) encountered when trying to use deepseek lite v2 with quantized K cache. Note: I compile my IK_Llama with GGML_CUDA_F16.
To fix this, I added a cpy_blck_q_f16 function devised by comparing the cpy_blck_q8_0_f32 and cpy_blck_q8_0_f16, and transposing the difference for the other legacy quants on the basis of the cpy_blck_q_f32 function. A "rule of three" of sorts.
Perplexity test and inference now works consistantly on -ctk q4_0 ; q4_1 ; q5_0 ; q5_1 in that scenario, with expected values and behavior.
Except on Q6_0, which sees its perplexity multiplied by 100. (I suspect the Cuda dequantize_q6_0 to be incompatible with this PR for some reason, but that's beyond what I can fix)
-ctk iq4_nl, which doesn't have yet a dequantize_iq4_nl function, is not usable that way for now.
* Fixing Gigachat support
* Gigachat: CUDA FA (needs 192 x 192 for MLA = 3)
* Gigachat: CPU FA (needs 192 x 192 for MLA = 3)
---------
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
* Fix q5_0_r4
The issue waqs in the tail part. As almost all models have tensor
rows that are multiple of 128, that part was never triggered in testing.
But ithe gpt-oss models have an embedding size of 2880, so we end
up there and trigger the bug.
* Fix q6_0_r4
Same fix as q5_0_r4
* Fix q4_0_r8
* Fix q5_0_r4 and q6_0_r4 also on Zen4
* Fix q4_0_r8 also on Zen4
---------
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
* Use new-new-mma also for MLA=3, and use mask bounds
This gives us ~25% better PP at 32k tokens compared to main
* This seems better
---------
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
* Fuse concat and copy into K cache
* Avoid ggml_cont() when n_token = 1
Combined effect: about +2% in TG performance with full GPU offload
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
This commit enables IQK quantization operations on ARM-based systems,
specifically tested on NVIDIA DGX Spark with GB10 Grace Blackwell.
Changes:
- Enable IQK_IMPLEMENT macro for ARM NEON operations
- Add arm_neon.h header include for ARM SIMD intrinsics
- Fix compilation errors related to missing NEON types and functions
Build requirements for ARM:
cmake .. -DGGML_CUDA=ON \
-DCMAKE_CXX_FLAGS="-march=armv8.2-a+dotprod+fp16" \
-DCMAKE_C_FLAGS="-march=armv8.2-a+dotprod+fp16"
Tested on:
- Platform: NVIDIA DGX Spark (aarch64)
- CPU: GB10 Grace Blackwell Superchip
- Memory: 128GB unified memory
Fixes build errors:
- 'float32x4_t' does not name a type
- 'vld1q_f32' was not declared in this scope
- 'v_expf' was not declared in this scope
- Missing FP16 NEON intrinsics
* Use mmq_id in mul_mat_id
* Better
* Also use it in the fused up+gate op
* Better -no-fmoe TG on CUDA
Still much slower than -fmoe, but abot 20-25% faster than what
we had before.
---------
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
* Introducing rope cache
When computing RoPE, the rotation angles in each layer
are exactly the same, and only depend on the token positions
(and other constant, model dependent parameters).
So, I wonder, why don't we compute the angles just once
and then reuse for the Q and K RoPE in each layer?
This commit does it as a POC on the CPU, and uses it in
the Qwen3-MoE compute graph.
* cuda: neox works
* WIP
* rope_cache: norm works
* Fused rope+rope
* Fused rope+rope (norm)
* Fused rms+rms+rope+rope (neox) - not working
* WIP
* Also qwen3
* Add command line arg to disable rope cache
* Disable RoPE cache if rope type is not neox or norm
* Add missing break after merge with main
* Fused fused_rms+fused_rms+rope+rope (with -mqkv)
* Fused fused_rms+fused_rms+rope+rope (without -mqkv)
---------
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
* Biased mmvq: minor optimization
* Fusing Q and K rms_norm for TG on CUDA
* Remove commented out code
---------
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
* Don't use vector kernels if K or V are quantized
* Correctly determine if FA is supported
* Also wmma
* Minor
---------
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
* Fuse Q, K, V gemv+add
* More gemv+add fusing
* Faster copy when tensors are contiguous
Relevant for storing data into the KV cache. I see ~1% speedup
for fast models (Ling-mini-2.0, gpt-oss-20b, etc.)
* Cleanup
* Make sure the bias really is 1 row to use fusion
---------
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
* Args for MMVQ functions
* WIP
* Fused ffn_up*unary_op(ffn_gate) for MMVQ (no bias)
We see nearly 2% TG speedup for Ling-mini-2.0 and
about 1% for DeepSeek-Lite.
* Fused ffn_up*unary_op(ffn_gate) for MMVQ (with bias)
* Fusing also for iqk/trellis/repacked quants
* Fusing mmvq also in non-MoE up+gate
* Fuse mul_mat_id and add_id into a single kernel for mmvq
* Also iqk quants
* Split mmvq.cu and iqk_mmvq.cu into separate template instances
* Put iqk mmvq implementations into template instances
* Somehow I forgot to change the ggml_type in the legacy template calls
* Add disagnostics
* Disable assert
* Fix TG fused up*nary(gate) when down cannot be fused
The wrong memory buffer got used in that case
---------
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
* Adding fused mul+multi_add + CPU implementation
* fused mul+multi_add: command line argument to disable it
* Faster tensor name formatting
We gain ~1% for Ling-mini-2.0 when running on CUDA.
---------
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
* Adding fused mul+multi_add + CPU implementation
* fused mul+multi_add: CUDA
* fused mul+multi_add: command line argument to disable it
---------
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
* Fuse add+add+fused_rms
* Try this
* Macro to easily enable/disable fusion
* Various:
* Check that all tensors involved are on the same device before applying fusion
* Fuse sigmoid+scale+sum_rows+div
* Fix the fused bailingmoe2 experts selection
The issue there was that the bias was not per row, but per
expert group, so only the first n_per_group biases were used
for al experts.
---------
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
* Combine all calls to llm_build_norm to a single line
so more easily check what kind of arguments are being passed
by simply using grep.
* Combine add + fused_rms_norm
For many models this happens at each layer: the result of the
layer is added to the ayer input, which then becomes the input
to the next layer, which then is typically normalized via
fused_rms_norm.
---------
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
* Fuse sigmoid+add+grouped_topk+get_rows (CPU)
* Fix CPU + CUDA
but CUDA is somehow not 100% correct as I get a slightly different
PPL (lower!)
* Minor
* Fuse sigmoid+add+topk+get_rows (CUDA)
* Fuse sigmoid+add+topk+get_rows (CPU)
* Fuse topk+view+get_rows+reshape+softmax (CPU)
* Fuse topk+view+get_rows+reshape+softmax (CUDA)
* cpu: turn off the openai topk fusing for now
Something is not right and I don't see the bug.
On the CPU one doesn't gain much if anything, so not a big loss.
* Also fuse sum_rows and div
---------
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
* Better argsort (CPU)
* Attemt at grouped topk
* This seems to do the trick for grouped experts routing
* Cleanup
* Trying to merge, something is not right
* Working merged grouped top_k (CPU)
* Add command line option to enable grouped expert routing
* Add grouped expert routing option to llama-bench
---------
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
* Add mtmd: the beginning
* Add mtmd: mtmd.cpp compiles
* Add mtmd: clip initialization compiles
* Add mtmd: clip.cpp compiles
* Add mtmd: builds successfully
* Add CPU implementation for GGML_OP_GLU
* Add CUDA implementation for GGML_OP_GLU
* Add CPU implementation for GGML_OP_CONV_2D and GGML_OP_CONV_2D_DW
* Add CUDA implementation for GGML_OP_CONV_2D and GGML_OP_CONV_2D_DW
* Add mtmd: refresh CPU rope
* Add mtmd: refresh CUDA rope
* Add mtmd: add Qwen2-VL
* Add mtmd: Qwen2.5-VL text seems to work with this change
* Add mtmd: fix swiglu
* Add mtmd: use LOG_TEE so generated tokens show up in terminal
* Add mtmd: do not attempt to load a GPU backend if none are available
* GLU, not GPU
* Fix typo
* Fix new/free mismatch
* LOG stuff
* Add mtmd: this fixes gibberish on second image
---------
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
* Avoid computing FA chunks where the mask is -infinity
* Avoid computing FA chunks where the mask is -infinity also for f16/bf16
---------
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
* Quick attempt to fuse the Q, K, V GEMMs
Doesn't do much on the CPU
* Doesn't do much on the GPU either
* Use llm_build_mul_mat_qkv
* This is not needed
* Revert timing on committed by mistake
---------
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
* Offload only activated experts
* This seems to do the trick for -fmoe
* Do not recalculate activated expers for fused up/gate
* Log out of bounds access details
* Add a command line argument
---------
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
* Bounds for flash attention
* Add n_swa to FA parameters
* Fix it
* This seems very slightly better
* Using vec kernel when we have SWA
* Need also this
* f32 vec kernel
* This is slightly better
---------
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
* Fused up+gate+unary for regular (not MoE) FFN - CPU
* WIP CUDA
* Seems to be working on CUDA
For a dense model we get 2-3% speedup for PP and ~0.6% for TG.
* Add command line option
This time the option is ON by default, and one needs to turn it
off via -no-fug or --no-fused-up-gate
---------
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
* Skip the row id computation for the ffn_down op
Sadly, almost negligible performance gain.
* Also this doesn't do much
* Also this barely moves the needle
* This is slightly better
---------
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
* Check for NaNs while loading the model.
* Also tell which experts have NaNs.
* Add command line option to validate quants
* Add checks for more quantization types
* Add checks for more quantizagtion types
---------
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
* This fixes confusion around Q8_0 on AVX2
* This does it for iq4_nl, including FA
* This does it for iq4_nl on Zen4, but FA does not work
* Slightly more clear
* Adding forgotten q8_0_r8 to num_rows()
---------
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
* Use bperm trick for iq2_ks gemm -> 7% gain
* Use bperm trick for iq2_k gemm -> ~5% gain
* Use bperm trick for iq2_k_r4 gemm -> ~3% gain
* Use bperm trick for iq2_ks gemv -> ~7% gain
* Use bperm trick for iq2_k gemv -> ~3% gain
* Use bperm trick for iq2_k_r4 gemv -> ~7% gain
---------
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
* q8_k_r16: basics
* q8_k_r16: iq4_xs now uses q8_k_r16 on Zen4+
PP performance is about the same as using q8_k_r8 on the Ryzen-7950X,
so we expect nice gains on Zen5, and we don't need to wory about
using 2 different q8_k_r8 implementations for fancy SIMD.
* q8_k_r16: iq2_xxs now uses q8_k_r16 on Zen4+
* q8_k_r16: iq2_xs now uses q8_k_r16 on Zen4+
* q8_k_r16: iq2_s now uses q8_k_r16 on Zen4+
* q8_k_r16: iq3_xxs now uses q8_k_r16 on Zen4+
* q8_k_r16: iq3_s now uses q8_k_r16 on Zen4+
* q8_k_r16: iq1_s and iq1_m now uses q8_k_r16 on Zen4+
* q8_k_r16: q2_K and q3_K now uses q8_k_r16 on Zen4+
* q8_k_r16: iq2_ks and iq2_k now uses q8_k_r16 on Zen4+
* q8_k_r16: iq2_kl now uses q8_k_r16 on Zen4+
* q8_k_r16: iq3_ks and iq3_k now uses q8_k_r16 on Zen4+
* q8_k_r16: iq4_kss, iq4_ks, and iq4_k now use q8_k_r16 on Zen4+
* q8_k_r16: iq5_ks, iq5_k, and iq6_k now use q8_k_r16 on Zen4+
* Fix AVX2
* Just always set num_rows to 16
---------
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
* Use bperm trick for iq3_ks - 5% PP performance gain
* Use bperm trick for iq3_k -> 5% PP performance gain
* Use bperm trick for iq3_k -> 8% PP performance gain
* Use bperm trick for iq3_k_r4 gemv -> ~5% faster
* Use bperm trick for iq3_k gemv -> ~3% faster
* Use bperm trick for iq3_k gemv -> 4.5% gain
---------
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
* Use __byte_perm in get_int_from_table_16
* Use get_int_from_table_16 everywhere for 4-bit quants
---------
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
Q8_0 needs Q0_0_X4, but Q8_0_R8 needs Q8_2_X4.
So, if we decide to repack a Q8_0 MoE tensor to Q8_0_R8,
iqk_moe_fused_mul_unary fails because the activations were
prepared as Q0_0_X4, but we now need Q8_2_X4.
For now a simple fix: just take the slow path, do not repack.
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
* This does the trick for PP
* Compute mask bounds when creating the mask
* Set mask bounds for all supported SWA models
---------
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
* gmp-oss: common
* gpt-oss: attnetion sinks, swiglu_oai
* gpt-oss: WIP llama
Model loads and runs (CPU only), but PPL is much to high
(~1500 for 1st batch vs ~200 in mainline).
Is it because of SWA, because of vocab, or did I introduce a bug somewhere?
* gpt-oss: CPU seems to be working
It was the SWA thta was missing in the previous commit.
There are issues with EOG tokens, so this still needs to be added.
* CUDA: ADD_ID
Just a copy from mainline
* gpt-oss: Seems to be working on CUDA
* gpt-oss: add sinks to the attn-vec kernels
* CUDA: add head size of 64 to new mma
Haven't turned it on yet, but observe slightly better PP and slightly
worse TG performance with that.
* gpt-oss: add ability to use -fmoe (only CUDA for now)
* Move row sums to the write place
* Add sinks to iqk flash attention
* gpt_oss: Implement -fmoe on the CPU
* Simdify swiglu_oai
Turning it off for now as performance becomes more variable,
so perhaps I'm running into thermal trottling imore often
because of making the CPU work too hard.
* llama: factor out model loader
* Builds successfully
* It runs, but mmap does not work
* Fix llama_mmap so mmap works
* Minor
* Fix CUDA after latest changes
* Attempt to use CUDA graphs with MoE models - not working
* CUDA graphs WIP - still not working
* CUDA graphs - seems to be working
Likely not all MLA variants are working.
I no longer remember why I added the q8_0 cpy that
transposes the tensor, but if really needed, this is now
missing. Also missing is q6_0.
* Make q8_0 cache work for DeepSeek models with CUDA graphs
* cuda: cpy for q6_0
* Fix llama_mmap on non-Linux platforms
* Adding forgotten file
* Iterating on Windows build failures
* cuda: re-add q8_0 -> q8_0 transpose
so mla = 2 can be used with CUDA graphs and q8_0 cache.
* Disable graphs without -fmoe
* Minor
* Turn graphs on by default
---------
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
* mxfp4: basics
* mxfp4: Zen4 GEMM
* mxfp4: repacked GEMM (AVX2/Zen4)
* mxfp4: AVX2 GEMM
* mxfp4: NEON GEMM
* mxfp4: repacked GEMM (NEON)
* mxfp4: Metal
* Fix quantized K cache without FA (#680)
* Prevent assert with quantized K cache and no FA
* Fix MMQ when running with quantized K cache without FA
---------
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
* Fix for Deepseek r1 parsing (#676)
* Implement function calling / tools for ik_llama.cpp for Kimi K2
* Implement basic tool choice
* Backport llama.cpp tool calls support
* Enhance function calls with improved chat parser and string utilities
- Add new chat.h/chat.cpp and chat-parser.h/chat-parser.cpp for better chat handling
- Improve function calls parsing with fallback to llama.cpp builder pattern
- Add string utility functions (starts_with, ends_with, find_partial_stop)
- Update README with function calls testing instructions
- Enhance Kimi K2 parser and function calls documentation
- Add comprehensive test suite for function calls
- Update CMakeLists.txt and Makefile for new components
* Enhance function calling with unified streaming and parser improvements
- Fix streaming content cleanup to prevent function syntax in output
- Unify content extraction patterns with llama.cpp approach
- Improve Kimi K2 parser robustness and partial content handling
- Add comprehensive test coverage for function call scenarios
- Optimize chat message parsing and diff computation
* Replace hardcoded values in kimi_k2_parser.hpp with named constants
- Add compile-time constants for all token format markers
- Add compile-time constants for XML format markers
- Add compile-time constants for simple format patterns
- Replace all hardcoded string literals with named constants
- Use compile-time length calculation to avoid manual counting
- Improve maintainability and reduce magic numbers throughout parser
* Fix duplicate common_chat_parse definition
- Remove duplicate implementation from chat-parser.cpp
- Keep single implementation in chat.cpp following llama.cpp patterns
- Resolves linker error: multiple definition of common_chat_parse
* Fix JSON assertion failure in function call parsing
- Add proper validation that 'function' field is an object before accessing nested keys
- Handle missing 'arguments' field gracefully with default "{}"
- Prevents crash when parsing malformed tool call JSON structures
* Add comprehensive Qwen3 XML tool calling support with unit tests
- Implement Qwen3 XML parser with <tool_call>{"name": "func", "arguments": {...}}</tool_call> format
- Add model detection and routing for Qwen3 vs Kimi-K2 formats
- Create 8 comprehensive unit tests covering parsing, streaming, error handling
- Fix token format cleaning bug in kimi_k2_parser.hpp processing order
- Remove progressive parsing code and related utilities
- Add tool injection support for Qwen3 format in server utils
* Add DeepSeek R1 function calling support with comprehensive unit tests
- Implement complete DeepSeek R1 tool call parsing in common_chat_parser.cpp
- Add DeepSeek R1 model detection and tool injection in deepseek_r1_tools.hpp
- Update function_calls.hpp with DeepSeek R1 integration and content extraction
- Update documentation to reflect support for Kimi-K2, Qwen3, and DeepSeek R1 models
- Add comprehensive unit tests for DeepSeek R1 reasoning, tool calls, and integration
- Port exact implementation patterns from original llama.cpp for compatibility
Key features:
- Native DeepSeek R1 format: <｜tool▁calls▁begin｜>function<｜tool▁sep｜>name```json{}```<｜tool▁call▁end｜><｜tool▁calls▁end｜>
- Reasoning content extraction from <think>...</think> tags
- Multiple tool calls support with separate call blocks
- Model detection for deepseek-r1, deepseek_r1 naming patterns
- Integration with incremental parsing and streaming support
* Add partial parsing support for JSON and regex
- json-partial.h/cpp: JSON partial parsing functionality
- regex-partial.h/cpp: Regex partial parsing functionality
* Add format_chat integration tests for Qwen3 tool injection
- Add test_qwen3_format_chat_integration() to validate tool injection pipeline
- Test tool injection conditions and system message enhancement
- Verify JSON formatting and anti-preamble instructions
- Add comprehensive test documentation
Tests confirm tool injection works correctly - conversational preamble
issue is not in ik_llama.cpp but likely in UI configuration.
* Fix Qwen3 tool call parsing - pass model name to parser
Server was not passing model name to parse_chat_message_incremental(),
causing Qwen3 to fall back to Kimi-K2 parser and return tool calls
as content instead of proper tool_calls array.
* Fix non-streaming path to use model-specific parsing
Non-streaming responses were hardcoded to use Kimi-K2 format,
causing Qwen3 XML tool calls to be returned as content instead
of proper tool_calls array. Now uses same model detection as
streaming path for consistency.
* Update Qwen3 function call handling in server and tests
- Enhanced server function call detection and response formatting
- Improved test coverage for Qwen3 tool call scenarios
- Refined XML parsing for better tool execution support
* Add DeepSeek-R1 function call parsing support
Implements comprehensive parsing for all 4 DeepSeek-R1 function call formats:
- Format 1: Standard function call syntax (already supported)
- Format 2: Alternative function call patterns (already supported)
- Format 3: Tools array format - function\n```json\n{"tools": [...]}
- Format 4: XML wrapped format - <tool_call>function</think>Name\n```json\n{...}```</tool_call>
Key changes:
- Added parse_deepseek_r1_tools_array() following original parse_prefixed_json_tool_call_array pattern
- Added parse_deepseek_r1_xml_wrapped() following Hermes-2-Pro XML wrapper patterns
- Integrated both parsers into exception handling chain for robust fallback
- Added comprehensive TDD test coverage for all formats
- Anonymized all confidential information while preserving functionality
Resolves tool_calls_count=0 issue where DeepSeek-R1 models generated valid tool calls
but server failed to parse them correctly.
* Update function_calls.md documentation for DeepSeek-R1 Format 4
- Added Format 4 (XML wrapped) documentation with examples
- Updated implementation notes with correct parser order (3→4→1→2)
- Marked all DeepSeek-R1 formats as working (July 2025 update)
- Updated test status for Format 3 and 4 as passing
- Added parse_deepseek_r1_xml_wrapped() function reference
- Corrected implementation file line numbers
* Fix merge conflict in test-function-calls.cpp
- Removed incomplete merge conflict marker from line 3027
- Ensured all tests compile and pass successfully
- All DeepSeek-R1 formats (1-4) working correctly
- All streaming and content cleaning tests passing
* Fix DeepSeek R1 parsing issue with responses wrapped in think tags
Restore missing consume_rest() call from working PR #648 implementation.
When responses don't contain tool calls, remaining content after reasoning
parsing must be preserved as displayable content.
Fixes issue where entire responses wrapped in <think> tags resulted in
empty content output.
* Implement proper reasoning handling following original llama.cpp patterns
- Add missing reasoning_format and reasoning_in_content fields to common_chat_syntax
- Update try_parse_reasoning to match original llama.cpp logic exactly
- Add TDD test case with reasoning_in_content=true for DeepSeek R1
- Following TDD: test should now pass with proper syntax configuration
Based on original llama.cpp implementation patterns.
* TDD SUCCESS: Fix DeepSeek R1 thinking tag termination issue
✅ Test passes with reasoning_in_content=true configuration
- Content properly preserved: '<think>content</think>' displays fully
- Reasoning field empty as expected
- Following TDD: test-first approach validates the fix
Next: Update server to automatically apply this configuration.
* Complete server integration fix for DeepSeek R1 thinking tag termination
- Server now automatically sets reasoning_in_content=true for DeepSeek R1 models
- Fixes issue where responses wrapped in <think> tags appear empty to users
* Add TDD test case for DeepSeek R1 thinking tag termination issue
- Test reproduces the exact failure scenario reported by user
- Validates that reasoning_in_content=true fixes the issue
- Demonstrates empty content problem and working solution
* Add remaining TDD test changes for DeepSeek R1 thinking tag fix
* Add debug output after upstream merge
* Remove temporary benchmark and debug files
- Remove tests/benchmark-progressive-parsing.cpp (development tool, not part of core functionality)
- Remove tests/reproduce_bug.sh (debugging script, not needed for PR)
* Port cpu moe options from mainline (#672)
* Port cpu moe options from mainline
* Use strdup and int32_t to follow coding guidelines
* maxfp4: CUDA dequantize
* mxfp4: CUDA GEMV
* mxfp4: CUDA MMQ
* mxfp4: minor CUDA tweaks
---------
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
Co-authored-by: Anton Sokolchenko <wsevendays@gmail.com>
Co-authored-by: Parsa <61601745+TheLegendOfKitty@users.noreply.github.com>
* Prevent assert with quantized K cache and no FA
* Fix MMQ when running with quantized K cache without FA
---------
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
* iq1_kt: basics
* iq1_kt: CUDA dequantize
Testing with LlaMA-3.1-8B-Instruct, we get almost the same PPL
as iq2_xxs, so about 0.2 bpw fewer bits for the same quality.
* iq1_kt: CUDA MMQ
* iq1_kt: CUDA MMVQ
* iq1_kt: AVX2 GEMM/GEMV
* iq1_kt: convert/repack to q8_0_r8 (AVX2)
* iq1_kt: slightly faster GEMV
18.6 t/s -> 19.4 t/s
* iq1_kt: NEON GEMM/GEMV
Pathetic as usual
* iq1_kt: slightly faster NEON - still pathetic
* iq1_kt: tiny bit better GEMV on NEON
* iq1_kt: convert/repack to q8_0_r8 (NEON)
* iq1_kt: very slightly faster convert/repack to q8_0_r8 on NEON
* Adding frgotten file
* iq1_kt: add to constants.py
---------
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
Kawrakow
firecoperana
firecoperana
abc-nix
Djip007
Nexes the Elder
Lennart Lopin
Thireus ☠
usrlocalben