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.
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 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.
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.
* CANN: Use smart pointers to manage ACL objects
Previously, ACL objects were managed via manual destruction, which
led to multiple memory-leak issues during runtime. This patch replaces
manual memory management with smart pointers so that ACL objects
are properly released and ownership is clearly defined.
Note that the ownership of an ACL object belongs to the function
that creates it. Other internal functions should operate on these ACL
objects using raw pointers to avoid unintended ownership transfers.
Additionally, since aclTensorList automatically frees its contained
aclTensor objects, any aclTensor added to a tensor list must release
ownership to avoid double free operations.
This PR also removes the asynchronous task submission mechanism.
Due to changes in recent CANN versions, tiling time has significantly
decreased. Even with a dual-thread submission model, the dispatch
overhead still falls on the critical path, making async submission
less beneficial. Moreover, aclGraph support provides a much better
path to reducing operator dispatch latency.
* CANN: resolve review comments
* vulkan: add LOG operation support for F32 and F16
Part of #14909.
* vulkan: Fix LOG operation types
* docs: Update operation support documentation for Vulkan LOG operation
* vulkan: fix log_f16 shader
* docs: restore missing LOG test cases and regenerate ops.md
* SYCL: add generic unary op implementation for multiple ops (ABS/SGN/…); unify non-contiguous access
* SYCL: update documentation and sycl.csv to reflect new unary op support
* update ops.md after syncing SYCL.csv changes
* Fix SYCL.csv merge conflict
* Update ops.md after fixing SYCL.csv conflicts
* Fix SYCL.csv tail after merge conflict and regenerate ops.md
* Fix line endings and final newline in SYCL.csv
* Remove TOPK_MOE entries from SYCL.csv as requested
* Update ops.md after removing TOPK_MOE from SYCL.csv
* Regenerated SYCL.csv and synced ops.md with upstream
* Update ops.md using create_ops_docs.py
* webui: add OAI-Compat Harmony tool-call live streaming visualization and persistence in chat UI
- Purely visual and diagnostic change, no effect on model context, prompt
construction, or inference behavior
- Captured assistant tool call payloads during streaming and non-streaming
completions, and persisted them in chat state and storage for downstream use
- Exposed parsed tool call labels beneath the assistant's model info line
with graceful fallback when parsing fails
- Added tool call badges beneath assistant responses that expose JSON tooltips
and copy their payloads when clicked, matching the existing model badge styling
- Added a user-facing setting to toggle tool call visibility to the Developer
settings section directly under the model selector option
* webui: remove scroll listener causing unnecessary layout updates (model selector)
* Update tools/server/webui/src/lib/components/app/chat/ChatMessages/ChatMessageAssistant.svelte
Co-authored-by: Aleksander Grygier <aleksander.grygier@gmail.com>
* Update tools/server/webui/src/lib/components/app/chat/ChatMessages/ChatMessageAssistant.svelte
Co-authored-by: Aleksander Grygier <aleksander.grygier@gmail.com>
* chore: npm run format & update webui build output
* chore: update webui build output
---------
Co-authored-by: Aleksander Grygier <aleksander.grygier@gmail.com>
* vulkan: change graph_compute to be async and enable get_tensor_async
This allows some additional CPU/GPU overlap for large pp workloads. Also seems
to help a bit for token gen, maybe getting rid of a small bubble between
graph_compute and get_tensor.
Async set and copy functions seem to be very rarely used, so I didn't enable
them because I didn't have a good way to test them.
The async commands need to be ordered against each other, so put them all on
the compute queue. The non-async commands still use the transfer queue.
The fence for graph_compute/get_tensor_async is submitted and waited on in
ggml_vk_synchronize.
* fix thread safety errors
* teardown context cleanly
* Handle async read to non-pinned dst
* fix : Dangling pointer for non-empty trigger words in llama_sampler_init_grammar_impl (#17047)
* Replace 'static' workaround, with keeping variable in scope for longer
* Create std::array directly and pass into llama_grammar_init_impl
* Add back the trigger pattern
* Missed array include
* ggml-cpu: handle 3d tensors in repack mul_mat
* Removed unnecessary branch, removed need for <algorithm>
* Fixed dst_ptr pointer in chunk + clang_format
* GGML_ASSERT to check wdata within bounds
* Accidental ggml.h inclusion
* Improved GGML_ASSERT on wdata boundaries
* Address performance regression in Qwen and llama.cpp due to chunking
Daniel Bevenius
Oliver Simons
Georgi Gerganov
Adrien Gallouët
hipudding
Pavels Zaicenkovs
Ruben Ortlam
Eve
lhez
shaofeiqi
shani-f
Aleksander Grygier
Pascal
Sigbjørn Skjæret
Jeff Bolz
Ankur Verma
Giuseppe Scrivano
Xuan-Son Nguyen
Bartowski
Marek Hradil jr.
ixgbe
Alberto Cabrera Pérez
Piotr Wilkin (ilintar)