chore: restore symmetry of worker start/shutdown
chore: fix return type of cal_max_tokens
chore: type some more return values
fix: free resources before re-claiming
Signed-off-by: ixlmar <206748156+ixlmar@users.noreply.github.com>
* refactor: Copy sequence lengths once in decoder setup
Signed-off-by: Robin Kobus <19427718+Funatiq@users.noreply.github.com>
* refactor: Update DecoderInputBuffers to remove duplicated buffers
- Renamed and reorganized buffer variables in decoderBuffers.h and decoderBuffers.cpp for better readability.
- Adjusted references in generateRequestOptions.cpp to align with the new buffer structure.
Signed-off-by: Robin Kobus <19427718+Funatiq@users.noreply.github.com>
* refactor: Move getEmbeddingBias to anonymous namespace
Signed-off-by: Robin Kobus <19427718+Funatiq@users.noreply.github.com>
* refactor: Filter context requests
Signed-off-by: Robin Kobus <19427718+Funatiq@users.noreply.github.com>
* refactor: GenerateRequestOptions using more fine-grained functions
- Added a new method `createDecoderRequests` to encapsulate the logic for creating decoder requests from finished context requests.
- Updated the `operator()` method to utilize the new method, improving code clarity and maintainability.
Signed-off-by: Robin Kobus <19427718+Funatiq@users.noreply.github.com>
Signed-off-by: Robin Kobus <19427718+Funatiq@users.noreply.github.com>
* refactor: Update TRTLLMDecoder
- Updated the `generate_request_options` call.
- Updated the `make_decoding_batch_input_output` call.
Signed-off-by: Robin Kobus <19427718+Funatiq@users.noreply.github.com>
* refactor: Remove const where we modify input buffers
- Changed `DecoderInputBuffers` parameters from const references to non-const references in multiple functions to allow modifications.
- Updated related function calls to ensure compatibility with the new parameter types.
Signed-off-by: Robin Kobus <19427718+Funatiq@users.noreply.github.com>
* fixup! refactor: Copy sequence lengths once in decoder setup
Signed-off-by: Robin Kobus <19427718+Funatiq@users.noreply.github.com>
---------
Signed-off-by: Robin Kobus <19427718+Funatiq@users.noreply.github.com>
* add docstring to summarize current rope support
Signed-off-by: Frida Hou <201670829+Fridah-nv@users.noreply.github.com>
* minor: replace call_method, adjust inserting order of cos_sin_cache calculation node
Signed-off-by: Frida Hou <201670829+Fridah-nv@users.noreply.github.com>
* add unit test for triton rope and ds rope
Signed-off-by: Frida Hou <201670829+Fridah-nv@users.noreply.github.com>
* update rope matcher to match DS RoPE, add custom op for reference, add unit test case
Signed-off-by: Frida Hou <201670829+Fridah-nv@users.noreply.github.com>
* cache cos[pos_idx].unsqueeze and sin[pos_idxs].unsqueeze
Signed-off-by: Frida Hou <201670829+Fridah-nv@users.noreply.github.com>
* minor doc update
Signed-off-by: Frida Hou <201670829+Fridah-nv@users.noreply.github.com>
* separate pattern matching and optimization for explicit and complex rope + minor updates
Signed-off-by: Frida Hou <201670829+Fridah-nv@users.noreply.github.com>
* clean rope impl in repo
Signed-off-by: Frida Hou <201670829+Fridah-nv@users.noreply.github.com>
* replace fused_flattened_mla_with_cache's rope impl with torch_apply_rope_with_qk_interleaving, update unit test
Signed-off-by: Frida Hou <201670829+Fridah-nv@users.noreply.github.com>
* minor
Signed-off-by: Frida Hou <201670829+Fridah-nv@users.noreply.github.com>
* separate layout infer and transpose to a new transformation
Signed-off-by: Frida Hou <201670829+Fridah-nv@users.noreply.github.com>
* update rope_with_explicit_freqs and rope_with_input_interleaved to expose unsqueeze_dim and support match_rope_layout, add unit tests
Signed-off-by: Frida Hou <201670829+Fridah-nv@users.noreply.github.com>
* solve merge conflict in transform.py, need to fix optimize_rope with cuda graph capture
Signed-off-by: Frida Hou <201670829+Fridah-nv@users.noreply.github.com>
* minor clean up after rebase
Signed-off-by: Ubuntu <201670829+Fridah-nv@users.noreply.github.com>
* fix pre-commit
Signed-off-by: Frida Hou <201670829+Fridah-nv@users.noreply.github.com>
* support map to bnsd layout and infer unsqueeze_dim from op
Signed-off-by: Frida Hou <201670829+Fridah-nv@users.noreply.github.com>
* fix cos/sin not the same across prompts in the same batch issue when mapping to flashinfer op
Signed-off-by: Frida Hou <201670829+Fridah-nv@users.noreply.github.com>
* fix for unit test
Signed-off-by: Frida Hou <201670829+Fridah-nv@users.noreply.github.com>
* fix custom op input/output node ordering issue for DeepSeek V3 rope
Signed-off-by: Frida Hou <201670829+Fridah-nv@users.noreply.github.com>
* clean code
Signed-off-by: Frida Hou <201670829+Fridah-nv@users.noreply.github.com>
* minor
Signed-off-by: Frida Hou <201670829+Fridah-nv@users.noreply.github.com>
* move flattening of cos_sin_cache to the graph, update flashinfer op docstring and test
Signed-off-by: Frida Hou <201670829+Fridah-nv@users.noreply.github.com>
* debug transform unit test failure
Signed-off-by: Frida Hou <201670829+Fridah-nv@users.noreply.github.com>
---------
Signed-off-by: Frida Hou <201670829+Fridah-nv@users.noreply.github.com>
Signed-off-by: Ubuntu <201670829+Fridah-nv@users.noreply.github.com>
Signed-off-by: Fridah-nv <201670829+Fridah-nv@users.noreply.github.com>
* update matcher to match expert compute first, then extract other args with LCA
Signed-off-by: Frida Hou <201670829+Fridah-nv@users.noreply.github.com>
* support 3D and 2D input in torch.ops.moe.trtllm_fused_moe
Signed-off-by: Frida Hou <201670829+Fridah-nv@users.noreply.github.com>
* update custom ops to support 3D and 2D inputs
Signed-off-by: Ubuntu <201670829+Fridah-nv@users.noreply.github.com>
* update deepseek patch
Signed-off-by: Ubuntu <201670829+Fridah-nv@users.noreply.github.com>
---------
Signed-off-by: Frida Hou <201670829+Fridah-nv@users.noreply.github.com>
* Add test case for kv memory estimation
* Dump running log into file and parse kv cache memory size from file
* Set bigger peak memory size for mixed percision case and test_ptp_quickstart_advanced_eagle3 case
* Revert change to usage of fraction
* use context manager to guard temp files
Signed-off-by: Hui Gao <huig@nvidia.com>
This PR adds a customized allreduce to TensorRT-LLM. The new allreduce is used for communication on PCIe-based GPUs via low-precision quantization, which can accelerate the PCIe allreduce process.
Signed-off-by: Hui Kang <hkang@nvidia.com>
Co-authored-by: Hui Kang <hkang@nvidia.com>
Support DeepSeek-R1 W4A8 on Hopper
Co-authored-by: Barry Kang <43644113+Barry-Delaney@users.noreply.github.com>
Co-authored-by: Jiang Shao <91270701+StudyingShao@users.noreply.github.com>
Signed-off-by: Barry Kang <43644113+Barry-Delaney@users.noreply.github.com>
* Add chunking to PyT heuristic.
Signed-off-by: Frank Di Natale <3429989+FrankD412@users.noreply.github.com>
* Cast tokens and batch size to ints.
Signed-off-by: Frank Di Natale <3429989+FrankD412@users.noreply.github.com>
---------
Signed-off-by: Frank Di Natale <3429989+FrankD412@users.noreply.github.com>
This issue is found for tp=ep=8 on the multi-node machine due to the inconsistent PP sizes.
* Reform the workspace allocation implementation to avoid the list-out-of-range issues.
* Disable min_latency_mode under the multi-node scenario to avoid the illegal memory access issue.
Signed-off-by: Yukun He <23156053+hyukn@users.noreply.github.com>
Prefetching safetensors files so that they are stored in the system file
cache. This significantly speeds up the model weight loading for the
very first run after entering the docker container.
This is beneficial because model weight loading is done layer-by-layer,
which means reading from the safetensors chunk-by-chunk, and that cannot
utilize the internet bandwidth very well, assuming that these files are
stored in some network drives. Instead, loading the whole files in bulk
can achieve higher internet bandwidth utilization.
When running with world_size>1, all ranks collaboratedly prefetch these
files.
In theory, we should add heuristics to decide whether to prefetch the
files or not, but that is beyond the scope of this commit.
For example, when the CPU memory is small, doing prefetching may result
in file cache thrashing, resulting in slower weight loading time.
Signed-off-by: Po-Han Huang <pohanh@nvidia.com>
* feat: Add heuristic for GroupRMSNorm kernel selection.
Implements a logistic regression model to dynamically select between:
- GroupRMSNormBaseKernel: Allocates warps proportional to sum of dimensions
(better SM occupancy in most cases)
- GroupRMSNormLargeBatch: Allocates warps proportional to max dimension
(better block scheduling in large batch scenarios)
Selection heuristic considers batch size, allocated warps, and scheduling
efficiency on the current GPU architecture. Models for Compute Capability
9.x and 10.x are trained base on nsys kernel runtime data.
The default kernel selection is the base kernel.
The python operator group_rms_norm will use the heuristic by default.
User can pick to use the base or large batch kernels as well.
Signed-off-by: Simeng Liu <simengl@nvidia.com>
* Address the comments.
Signed-off-by: Simeng Liu <simengl@nvidia.com>
---------
Signed-off-by: Simeng Liu <simengl@nvidia.com>
