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
```
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.
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.
* extract rotate_pairs logic from ggml_compute_forward_rope_f32
* templateify ggml_compute_forward_rope_f32 and _f16
* abort when rope type not supported, remove GLM from test-rope
* add imrope branch to switch
* add rope tests for perf
* Update ggml/src/ggml-cpu/ops.cpp
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* Update ggml/src/ggml-cpu/ops.cpp
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
---------
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* vulkan : implement upscale with bicubic interpolation
* cuda : implement upscale with bicubic interpolation
* tests : add ggml_interpolate with GGML_SCALE_MODE_BICUBIC to backend tests
* adapt OpenCL backend to not support the OP in that case so tests don't fail
* print scale mode & flags in test-backend-ops
This change combines the rms_norm+mul and rope+view+set_rows fusions to
allow fusing the whole sequence together. This comes up in Qwen3, Bailing,
and some other models.
* WIP
* added a cpy kernel specific to transposed tensor which uses smem to avoid uncoalesced access; test cases also added shwoing improved memory bandwidth
* added BF16 support
* more strict check to make sure src0 is a transpose
* reformulated to handle more complicated transpose cases
* bring back 2D transpose for higher performance
* allow build on windows
* tranpose copy more shapes
* minor tweak
* final clean up
* restore some test cases
* keep only the kernel for true tranposed case; updated with review suggestions
* make CI happy
* remove headers not needed
* reduced bank conflicts for fp16 and bf16
* add missing const*
* now bank conflicts free
* use padding instead of swizzling
---------
Co-authored-by: bssrdf <bssrdf@gmail.com>
* tests: fix segfault in moe-expert-reduce test in support mode and --show-coverage
* tests: init gf and filter out fusion tests for support mode
* tests: filter out fusion cases before calling eval_support
* tests: filter out fusion cases from show_test_coverage as well, fix lint
* 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>
* server : support unified context across slots
* cont : fix speculative decoding initialization
* context : fix n_ctx_per_seq computation
* server : purge slots one by one
* tests : add unified cache server tests
* llama : update per-seq context computation
* test-thread-safety : handle tiny training context of the input model
* server : fix server_tokens clear()
* server : use 4 slots + unified KV by default
* llama : add note about context size queries
* cont : update todos [no ci]
* context : do not cap the size of the context
* tests : adjust parameters to be CI friendlier
* context : add warning
This pattern appears in a lot of models, the rope operation is applied right
before storing into the KV cache (usually on the K tensor).
Add a path to some of the rope shaders that computes the destination address
based on the set_rows tensor. Compile variants of the shader with D_TYPE of
f16 (the usual KV cache type).
Add a src3 operand to ggml_vk_op_f32 - sometimes rope uses three srcs and needs
the fourth for the row indices.
Add fused_ops_write_mask to indicate which intermediate tensors need to write
their results to memory. Skipping writing the roped K value helps to allow more
nodes to run concurrently.
Add logic to ggml_vk_graph_optimize to make ROPE+VIEW+SET_ROWS consecutive. It
rarely starts out that way in the graph.
Add new backend tests.
* ggml : fix interpolate with align-corners and ne=1
* avoid division by zero if one of the spatial dimensions is 1
* cpu, cuda, opencl returned correct result anyway due to clamp
* vulkan didn't clamp for align-corners so results were broken
* fix clang warning
* SYCL: Add support for FLOOR,CEIL,ROUND and TRUNC unary operators
Clean up unrelated changes from previous commit
* Chore: remove empty lines and fix indentation
* Clean up: remove leftover blank lines and fix spacing
* chore: fix trailing whitespace and ensure final newline
* Cleanup: remove redundant declarations already defined in header
* Sync docs/ops.md with updated backend operation support
* docs: update ops.md after rebase
* docs: update ops.md - Vulkan supports SSM_CONV and SSM_SCAN
* opencl: add mm_q8_0_f32
* opencl: fix data loading for incomplete tile
* opencl: use q8_0 mm for larger matrix
* opencl: add some tests to cover the path
* optimise GGML_OP_SUM
* add non-contiguous tests by permuting the input
* change tests to require full contiguity of OP_SUM
* cuda : add check GGML_OP_SUM
---------
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* refactor: unify reasoning handling via backend reasoning_content, drop frontend tag parsing
- Updated the chat message component to surface backend-supplied reasoning via message.thinking while showing the raw assistant content without inline tag scrubbing
- Simplified chat streaming to append content chunks directly, stream reasoning into the message model, and persist any partial reasoning when generation stops
- Refactored the chat service SSE handler to rely on server-provided reasoning_content, removing legacy <think> parsing logic
- Refreshed Storybook data and streaming flows to populate the thinking field explicitly for static and streaming assistant messages
* refactor: implement streaming-aware universal reasoning parser
Remove the streaming mode limitation from --reasoning-format by refactoring
try_parse_reasoning() to handle incremental parsing of <think> tags across
all formats.
- Rework try_parse_reasoning() to track whitespace, partial tags, and
multiple reasoning segments, allowing proper separation of reasoning_content
and content in streaming mode
- Parse reasoning tags before tool call handling in content-only and Llama 3.x
formats to ensure inline <think> blocks are captured correctly
- Change default reasoning_format from 'auto' to 'deepseek' for consistent
behavior
- Add 'deepseek-legacy' option to preserve old inline behavior when needed
- Update CLI help and documentation to reflect streaming support
- Add parser tests for inline <think>...</think> segments
The parser now continues processing content after </think> closes instead of
stopping, enabling proper message.reasoning_content and message.content
separation in both streaming and non-streaming modes.
Fixes the issue where streaming responses would dump everything (including
post-thinking content) into reasoning_content while leaving content empty.
* refactor: address review feedback from allozaur
- Passed the assistant message content directly to ChatMessageAssistant to drop the redundant derived state in the chat message component
- Simplified chat streaming updates by removing unused partial-thinking handling and persisting partial responses straight from currentResponse
- Refreshed the ChatMessage stories to cover standard and reasoning scenarios without the old THINK-tag parsing examples
Co-authored-by: Aleksander Grygier <aleksander.grygier@gmail.com>
* refactor: restore forced reasoning prefix to pass test-chat ([chat] All tests passed)
- store the exact sequence seen on input when 'thinking_forced_open' enforces a reasoning block
- inject this prefix before the first accumulated segment in 'reasoning_content', then clear it to avoid duplication
- repeat the capture on every new 'start_think' detection to properly handle partial/streaming flows
* refactor: address review feedback from ngxson
* debug: say goodbye to curl -N, hello one-click raw stream
- adds a new checkbox in the WebUI to display raw LLM output without backend parsing or frontend Markdown rendering
* Update tools/server/webui/src/lib/components/app/chat/ChatMessages/ChatMessage.svelte
Co-authored-by: Aleksander Grygier <aleksander.grygier@gmail.com>
* webui: add Storybook example for raw LLM output and scope reasoning format toggle per story
- Added a Storybook example that showcases the chat message component in raw LLM output mode with the provided trace sample
- Updated every ChatMessage story to toggle the disableReasoningFormat setting so the raw-output rendering remains scoped to its own example
* npm run format
* chat-parser: address review feedback from ngxson
Co-authored-by: Xuan Son Nguyen <thichthat@gmail.com>
---------
Co-authored-by: Aleksander Grygier <aleksander.grygier@gmail.com>
Co-authored-by: Xuan Son Nguyen <thichthat@gmail.com>
* metal : pad K, V and Mask when needed
* cont : simplify
* cuda : add TODO about KV padding requirement
* metal : add comments
* metal : remove mask padding requirement
* tests : add -INF blocks to the KQ mask in the FA tests
* cont : bump -INF block size to 64
Co-authored-by: Jeff Bolz <jbolz@nvidia.com>
* ggml : prevent division by zero in FA CPU op
---------
Co-authored-by: Jeff Bolz <jbolz@nvidia.com>
* fix: Fix duplicate fake image before token on first slice
Branch: GraniteDoclingStopping
Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
* fix: Use double-newline before overview image
Branch: GraniteDoclingStopping
Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
* fix: Remove incorrect newline at the end of granite chat template gen prompt
There should not be one, even for the language models.
Branch: GraniteDoclingStopping
Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
* tests: Remove bad newline from granite chat template test (legacy)
Branch: GraniteDoclingStopping
Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
---------
Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
* feat: added a dedicated Magistral chat format that preserves [THINK] spans, parses reasoning before tool calls
* feat: new flow in the chat template test suite for Magistral
Oliver Simons
Daniel Bevenius
Georgi Gerganov
Jeff Bolz
Piotr Wilkin (ilintar)
Diego Devesa
duduta
Acly
Ruben Ortlam
Aman Gupta
bssrdf
Johannes Gäßler
Shagun Bera
JJJYmmm
Sigbjørn Skjæret
Aldehir Rojas
Yuri Khrustalev
leejet
safranowith
Ilia Ilmer
Olivier Chafik
lhez
Sam/Samuel
Pascal
Gabe Goodhart