* feat: trtllm-gen fp4 GEMM
Signed-off-by: Dom Brown <3886319+DomBrown@users.noreply.github.com>
* Clean up
Signed-off-by: Dom Brown <3886319+DomBrown@users.noreply.github.com>
* Remove incorrect header
Signed-off-by: Dom Brown <3886319+DomBrown@users.noreply.github.com>
* Reviewer comment
Signed-off-by: Dom Brown <3886319+DomBrown@users.noreply.github.com>
---------
Signed-off-by: Dom Brown <3886319+DomBrown@users.noreply.github.com>
* Instead of allocating UserBuffers at beginning of runtime, UB buffers
are now managed with global allocator. The allocator will dynamically
assign free UB buffer or allocate new buffer for torch tensor. It makes
userbuffers easier to use.
* In common usecase, the Userbuffers will be allocated correctly during
warm up stage. There is no dynamic allocation during inference.
* UB fusion pattern is rewroten using the new UB Allocator. It contains
following passes:
1. Fuse Quant with allreduce, replace with UB impl, and insert a
copy_to_userbuffers. Currently the normal allreduce still does not
support FP8 quant. So this need to be done in UB pass
2. Convert all supported allreduce with UB and insert copy_to_userbuffers.
3. Fuse op before ar with the copy_to_userbuffers. So the op directly
writes to the userbuffer
4. Remove userbuffers finalize if the output is connect to another UB
allreduce.
Signed-off-by: Jin Li <59594262+liji-nv@users.noreply.github.com>
* Several optimizations and fixings on the Autotuner.
Signed-off-by: Yukun He <23156053+hyukn@users.noreply.github.com>
* Apply the new Python side Autotuner on current linear for nvFP4 data type.
Signed-off-by: Yukun He <23156053+hyukn@users.noreply.github.com>
* Apply the new Python side Autotuner on MoE op
* Remove routers from cache key to improve inference perf
* Prevent unnecessary code profiling. Use do_preparation keyword to select which part should be executed during before evaluating any tactic.
* Remove try-catch inside moe profiling process.
* Move default tactic -1 to 0 transforms in cpp runner.
* Revise relavant tests.
* Predefined the bucketizing strategy for fused_moe
Signed-off-by: Yukun He <23156053+hyukn@users.noreply.github.com>
* Add specific_profile support for AutoTuner to bypass the standard cache search process for perf optimization
* Add specific_profile for moe
* Add specific profile for linear
Signed-off-by: Yukun He <23156053+hyukn@users.noreply.github.com>
* Fixing and revising according to reviewer's suggestions.
Signed-off-by: Yukun He <23156053+hyukn@users.noreply.github.com>
* Use lru_cache for inference pref optimization.
* Revert gen_custom_cache_key feature
Signed-off-by: Yukun He <23156053+hyukn@users.noreply.github.com>
* Replace runner with runner id to achieve a serializable cache.
Signed-off-by: Yukun He <23156053+hyukn@users.noreply.github.com>
* Code clean up and minor fixings.
Signed-off-by: Yukun He <23156053+hyukn@users.noreply.github.com>
* Move all tunable runners and custom ops into torch_custom_ops.
Signed-off-by: Yukun He <23156053+hyukn@users.noreply.github.com>
* Treat min_latency_mode as a independent dynamic tensor. Modify get_valid_tactics to suit for it.
Signed-off-by: Yukun He <23156053+hyukn@users.noreply.github.com>
---------
Signed-off-by: Yukun He <23156053+hyukn@users.noreply.github.com>
* feat: Add option to run disaggregated serving without ctx servers, to benchmark gen only
Signed-off-by: Patrice Castonguay <55748270+pcastonguay@users.noreply.github.com>
* Fixing comment in sanity check
Signed-off-by: Patrice Castonguay <55748270+pcastonguay@users.noreply.github.com>
---------
Signed-off-by: Patrice Castonguay <55748270+pcastonguay@users.noreply.github.com>
* fp8 kv + bf16 ctx MLA + fp8 gen MLA
Use BF16 for context MLA.
mFP8GenerationMLA and mFP8ContextFMHA shouldn't be enabled together.
Allow mSM==90 for mFP8GenerationMLA==true.
For FMHA, dataTypeKv should be FP8.
For FP8 MLA generation, the output is still in BF16.
Refine debug info for FMHA kernel metadata.
Use inputType, outputType, SM together to hash kernel list.
Add FP8 MLA generation FMHA kernel.
Special WAR of NUM_COMPUTE_GROUPS for MLA generation kernel.
Separate the implementation of fused_multihead_attention_v2.h to CPP and print some debug info if checkIfKernelExist fails.
Refine debug info in fused_multihead_attention_v2.cpp
Correct FP8 MLA metadata.
New kernel provided by Yuxin, which outputs BF16.
smem size is not set correctly, which will lead to illegal mem access.
Yuxin fixed the error in FMHA MLA kernel: previously the BF16 isn't correctly written: some parts are repeatedly written, while some others are untouched.
There are two bmm1 scales that should be set correctly.
New kernel generated by Yuxin.
Modificatiosn to common/attentionOp for FP8 MLA on Hopper using FMHA.
Not necessary. If mFP8GenerationMLA, is_fp8_out is false, so mFP8ContextFMHA is false.
Skip a check in fmhaDispatcher.
Modifications in fmhaRunner:
- Debug dump.
- if (!isFP8GenerationMLA) skips a lot of flag setting.
- TMA descriptor modification for qo (by Yuxin).
Cleanup debug output.
Clean up o tma descriptor modifications.
Signed-off-by: Bo Li <bobboli0202@gmail.com>
* Resolve conflicts.
Signed-off-by: Bo Li <bobboli0202@gmail.com>
* Apply the patch of FP8 FlashMLA and resolve conflicts.
Signed-off-by: Bo Li <bobboli0202@gmail.com>
* Fix compilation error.
Signed-off-by: Bo Li <bobboli0202@gmail.com>
* Fix compile error.
Signed-off-by: Bo Li <bobboli0202@gmail.com>
* pick blackwell support
Signed-off-by: Dylan Chen <191843203+DylanChen-NV@users.noreply.github.com>
* Add copyright notice to fused_multihead_attention_v2.cpp.
Signed-off-by: Bo Li <bobboli0202@gmail.com>
* Add license.
Signed-off-by: Bo Li <bobboli0202@gmail.com>
* Add missing license.
Signed-off-by: Bo Li <bobboli0202@gmail.com>
* Exclude building flashMLA kernels under sm90.
Signed-off-by: Bo Li <bobboli0202@gmail.com>
* Revert "Exclude building flashMLA kernels under sm90."
This reverts commit f0c859d459.
Signed-off-by: Bo Li <bobboli0202@gmail.com>
* Use macro to skip compiling FlashMLA for non sm90 targets.
Signed-off-by: Bo Li <bobboli0202@gmail.com>
---------
Signed-off-by: Bo Li <bobboli0202@gmail.com>
Signed-off-by: Dylan Chen <191843203+DylanChen-NV@users.noreply.github.com>
Co-authored-by: Dylan Chen <ziqingc@nvidia.com>
Co-authored-by: Dylan Chen <191843203+DylanChen-NV@users.noreply.github.com>
Co-authored-by: QI JUN <22017000+QiJune@users.noreply.github.com>
* refactor: Expose DecoderState via bindings and integrate in TRTLLMDecoder
- Introduced a new `DecoderState` class in the C++ bindings, encapsulating key functionalities for managing decoding state.
- Adjusted the Python `TRTLLMDecoder` to access properties from `decoder_state`, ensuring consistency and clarity in the decoding process.
These changes streamline the decoder's architecture and enhance maintainability.
Signed-off-by: Robin Kobus <19427718+Funatiq@users.noreply.github.com>
* chore: Remove unused new_tokens from DecoderState bindings
Signed-off-by: Robin Kobus <19427718+Funatiq@users.noreply.github.com>
---------
Signed-off-by: Robin Kobus <19427718+Funatiq@users.noreply.github.com>
* init trtllm attn no cache
Signed-off-by: Qixiang Lin <qixiangl@nvidia.com>
* fix: fix the seq_len issue and attn metadata prepare for qwen reward model test
fix: fix minor bugs after rebase
Signed-off-by: Qixiang Lin <qixiangl@nvidia.com>
* refactor: remove unnecessary debug logs and clean up commented code
refactor: update max_seq_len documentation and remove max_seq_len for decoder model contructor in PyTorchModelEngine
Signed-off-by: Qixiang Lin <qixiangl@nvidia.com>
* refactor: update calculate_ref_result function to accept tensor inputs and mask type, enhance test_attention_no_cache to support FULL and CAUSAL masks
Signed-off-by: Qixiang Lin <qixiangl@nvidia.com>
* refactor: remove unused BERT attention metadata conversion method and add type assertion for no cache attention in PyTorchModelEngine
Signed-off-by: Qixiang Lin <qixiangl@nvidia.com>
* refactor: remove use_kv_cache parameter from attention function and related classes, update documentation for KV cache handling
Signed-off-by: Qixiang Lin <qixiangl@nvidia.com>
* refactor: implement setAttentionMaskType method for better mask type handling and remove unused conversion function
Signed-off-by: Qixiang Lin <qixiangl@nvidia.com>
* refactor: streamline KV cache handling by replacing direct member access with useKVCache method and simplify token per block assignment
remove Debug code.
Signed-off-by: Qixiang Lin <qixiangl@nvidia.com>
* refactor: Resolve comments for Python code
Simplify no cache attention metadata preparation and streamline related attributes in TrtllmAttentionMetadata
Removed the private method for converting to no cache attention metadata and integrated its logic into the prepare method. Updated the test for BERT sequence classification to reflect these changes and ensure proper handling of attention metadata.
Signed-off-by: Qixiang Lin <qixiangl@nvidia.com>
* docs: Add is_dummy_attention field to attention metadata for simulation operations
Signed-off-by: Qixiang Lin <qixiangl@nvidia.com>
* refactor: add KVCacheParams to attention backend interface and import relevant metadata classes
Updated the attention backend interface to include KVCacheParams and imported TrtllmAttentionMetadata and VanillaAttentionMetadata in model_engine.py for enhanced functionality.
Signed-off-by: Qixiang Lin <qixiangl@nvidia.com>
* fix: fix rebase format issue
Signed-off-by: Qixiang Lin <qixiangl@nvidia.com>
* fix: extend attention mask type handling in MHARunnerFixedParams
Added support for additional attention mask types (BIDIRECTIONAL, BIDIRECTIONALGLM, BLOCKSPARSE) in the MHARunnerFixedParams structure to fix the mapping issue between ContextAttentionMaskType and AttentionMaskType
Signed-off-by: Qixiang Lin <qixiangl@nvidia.com>
* fix: enhance attention mask type handling in TllmGenFmhaRunnerParams
Updated the setAttentionMaskType method to include a switch-case structure for better handling of attention mask types, ensuring proper mapping and error handling for invalid types.
Signed-off-by: Qixiang Lin <qixiangl@nvidia.com>
---------
Signed-off-by: Qixiang Lin <qixiangl@nvidia.com>
* Reapply "refactor: Replace DecoderFinishedEvent with CudaEvent in decoder clas…" (#3183)
This reverts commit 75495730bc.
Signed-off-by: Robin Kobus <19427718+Funatiq@users.noreply.github.com>
* fixup! Reapply "refactor: Replace DecoderFinishedEvent with CudaEvent in decoder clas…" (#3183)
Signed-off-by: Robin Kobus <19427718+Funatiq@users.noreply.github.com>
---------
Signed-off-by: Robin Kobus <19427718+Funatiq@users.noreply.github.com>
- Updated the first dimension of additional output tensors to match mMaxNewTokens.
- Copy output of last context token to generation outputs.
- Adjusted the expected output size calculations in unit tests to reflect the correct maximum output length.
Signed-off-by: Robin Kobus <19427718+Funatiq@users.noreply.github.com>
Required to correctly support recent archs like 90a, ...
Fix issue #3173
Signed-off-by: William Tambellini <wtambellini@sdl.com>
Co-authored-by: Yuan Tong <13075180+tongyuantongyu@users.noreply.github.com>
* Update fp8 sf layout for blackwell and enable fp8 gemm e2e
* Add test case when m needs to be padded
* Better comment
Signed-off-by: Chang Liu <liuc@nvidia.com>
* Add TODO for fp8 quant kernel
Signed-off-by: Chang Liu <liuc@nvidia.com>
* Enable DCO check
Signed-off-by: Chang Liu <liuc@nvidia.com>
* Fix lint
---------
Signed-off-by: Chang Liu <liuc@nvidia.com>
* refactor: Simplifiy disableLookahead method
Signed-off-by: Robin Kobus <19427718+Funatiq@users.noreply.github.com>
* Update DecoderBuffers comments
Signed-off-by: Robin Kobus <19427718+Funatiq@users.noreply.github.com>
* refactor: Move numDecodingEngineTokens to DecoderState
This commit introduces new methods in the DecoderState class to manage the number of tokens for each request in a batch. The following changes were made:
- Added `getNumDecodingEngineTokens()` to retrieve the number of tokens for all requests.
- Added `getNumDecodingEngineTokens(SizeType32 batchIdx)` to get the token count for a specific request.
- Added `setNumDecodingEngineTokens(SizeType32 batchIdx, SizeType32 numTokens)` to set the token count for a specific request.
- Updated the setup method to initialize the token count vector based on the maximum batch size.
- Refactored the `CreateNewDecoderRequests` class to utilize the new token management methods, improving clarity and maintainability.
Signed-off-by: Robin Kobus <19427718+Funatiq@users.noreply.github.com>
* refactor: Improve shape variables in DecoderState
Signed-off-by: Robin Kobus <19427718+Funatiq@users.noreply.github.com>
---------
Signed-off-by: Robin Kobus <19427718+Funatiq@users.noreply.github.com>
Early exit if find_library() does not find any lib.
As today, the find_library_create_target() cmake macro
blindly continues even if the lib is not found,
adding LIB_PATH-NOTFOUND to the target and making the build
failing anyway later with non obvious reasons.
This change just early exits if the lib is simply not found with a
proper error message.
Fix github issue #3109
Signed-off-by: William Tambellini <wtambellini@sdl.com>
- Updated the `forwardAsync` method in `GptDecoderBatched` and `iGptDecoderBatched` to return `CudaEvent` instead of `DecoderFinishedEventPtr`, simplifying event handling.
- Removed the `DecoderFinishedEvent` class and its associated usage across various files, streamlining the codebase.
- Adjusted related methods and Python bindings to accommodate the new event structure, ensuring compatibility and maintaining functionality.
These changes enhance the clarity and efficiency of the decoding process in the batch manager.
Signed-off-by: Robin Kobus <19427718+Funatiq@users.noreply.github.com>
* refactor: Update gatherTree function to accept CUDA stream parameter
This commit modifies the gatherTree function signature to include a runtime::CudaStream parameter, enhancing flexibility in stream management. Additionally, it removes unnecessary buffer manager parameters and stream handling from the function, streamlining the code. The finalize method in GptDecoderBatched is also updated to reflect these changes, improving clarity and maintainability in the decoding process.
Signed-off-by: Robin Kobus <19427718+Funatiq@users.noreply.github.com>
* refactor: Update GptDecoderBatched finalize
This commit refactors the GptDecoderBatched class to improve method signatures and reduce code complexity:
- Modified finalize method to accept DecoderState as a parameter
- Updated method signatures to work with the new DecoderState approach
- Improved code organization and readability
The changes continue the ongoing refactoring to centralize decoder state management and simplify the decoder implementation.
Signed-off-by: Robin Kobus <19427718+Funatiq@users.noreply.github.com>
---------
Signed-off-by: Robin Kobus <19427718+Funatiq@users.noreply.github.com>
* chore: update cutlass to v3.8.0
Signed-off-by: Robin Kobus <19427718+Funatiq@users.noreply.github.com>
* refactor: update include directives for consistency and organization in weightOnlyBatchedGemv headers
Signed-off-by: Robin Kobus <19427718+Funatiq@users.noreply.github.com>
* Fix fpA_intB_gemm compilation
Signed-off-by: Robin Kobus <19427718+Funatiq@users.noreply.github.com>
---------
Signed-off-by: Robin Kobus <19427718+Funatiq@users.noreply.github.com>
* refactor: Remove ForwardType enum from GptDecoderBatched
- Remove ForwardType enum from GptDecoderBatched
- Simplify forwardDispatch and forwardDecoder methods
Signed-off-by: Robin Kobus <19427718+Funatiq@users.noreply.github.com>
* refactor: Remove forwardDecoder method from GptDecoderBatched
- Eliminate the forwardDecoder method to streamline the decoding process.
- Update forwardDispatch to directly call forwardAsync when input batch size is greater than zero.
Signed-off-by: Robin Kobus <19427718+Funatiq@users.noreply.github.com>
* refactor: Move event handling from forwardDispatch to forwardAsync
Signed-off-by: Robin Kobus <19427718+Funatiq@users.noreply.github.com>
---------
Signed-off-by: Robin Kobus <19427718+Funatiq@users.noreply.github.com>
