This merge request attempts to support more SWA KV cache functionality
inside the KV cache manager. Before this merge request, the KV cache for
sliding window attention (SWA) only holds "window size" number of blocks
and reuse them in a cyclic manner. We will not be able to utilize more
GPU memory with this design, leading to a limited max batch size
throughput. Additionally, we will not be able to support KV cache reuse
with this design.
In this MR, we change such behavior to let the manager write blocks in
a linear manner. With a linear block writing behavior, as the attention
window moves on, the out-of-window (OOW) blocks will be detached. Right
now for the sake of a correct feature first, we directly offload the
OOW block from the primary block pool (GPU memory) to the secondary
block pool (host memory). We will improve this in the future by
delegating the block movement to the eviction policy.
KV cache reuse for SWA is not developed in this merge request and will
be amended in a follow-up merge request.
Writing the blocks linearly, the maximum number of blocks allocated for
a sequence(`GenerationRequest`) is the "max sequence length" specified.
The `GenerationRequest` that stores the cache block bookkeeping
structure will now keep "max sequence length" tokens of blocks.
Given the above, main changes are (more context in the MR):
- Remove "cyclic" concept under the kv cache manager, such concept
originally guards the block reuse under kv cache manager.
- Add detach mechanism and have it under `KVCacheManager::addToken`.
Please note that detach is still guarded off for SWA when reuse
is enabled. A follow-up merge request will proceed to improve this.
- Enforce "max sequence length" to be a non-optional parameter to
the `KVCacheManager`/`BlockManager`
- Let all window size resource pool get identical proportion of memory
- Fix free memory calculation under `resource_manager.py`
Signed-off-by: eopXD <yuehtingc@nvidia.com>
Co-authored-by: Tomer Asida <tasida@nvidia.com>
Signed-off-by: Chang Liu (Enterprise Products) <9713593+chang-l@users.noreply.github.com>
Signed-off-by: Chang Liu <9713593+chang-l@users.noreply.github.com>
This MR is a preliminary MR for implementing the SWA reuse mechanism for
the kv cache manager. Please be aware that **no functional change is
intended** in this merge request. The purpose of the clean-up is to
decouple and remove existing functions for the up-coming SWA KV cache
reuse change to be more natural and easier to review.
Right now, (1) streamLLM, and (2) beam search with SWA, are broken. We
do not want to complicate the code base by stacking more features upon
something that does not work. This MR prunes out the logic and add
assertions so we can come back and re-support the broken feature and
remove the assertion.
Since streamLLM (sink attention) is broken now, assertion is added
under `KVCacheManager` ctor to guard for the value of
`mSinkBlockTokenLength` and `mSinkBubbleLength`. Compute logics relate
to it are pruned.
The beam search with SWA will still be broke when introducing the SWA
KV cache reuse. We will revisit this problem in the future.
On top of this, we should make an effort to update the [supporting matrix](https://github.com/NVIDIA/TensorRT-LLM/blob/feat/1.0_doc_dev/docs/source/1.0/features/feature-combination-matrix.md)
of the kv cache manager after merging the support of SWA KV cache reuse.
Changes are listed as following:
- Separate `KVCacheManager::updateToken` into `KVCacheManager::addToken`
and `KVCacheManager::removeToken`. The functionality should be
decoupled.
- Push utility `cacheSequenceBlockOffsets` and `cacheNewBlockOffset` from
`KVCacheManager` down to `WindowBlockManager`. `KVCacheManager`-exposed
functions should be real utilities that users of the structure can
leverage. Implementation-detailed function calls should not exist at
this level.
- Simplify "is shared last context block" logic under
`KVCacheManager::addSequence`.
Since no functional change is intended in this merge request, no test
case is added. Several comments are added for future test coverage
reminder.
For `LlmRequestTest.ParamTest`, `streaming=True` is commented out
because we guard sink attention with assertion now.
In `capacitySchedulerTest`, `addToken` action to `crossKVCacheManager`
is removed because in encoder-decoder model, generation tokens are
added only to the decoder and not to the encoder.
Signed-off-by: eopXD <yuehtingc@nvidia.com>
No functional change is intended in this MR.
`WindowBlockManager::mCachedBlocksRoot` is now who is responsible
for the bookkeeping of the `KVCacheBlock`, and the `mNextBlocks` is
now the actual hash map that fetches the block.
The `mEnableHashKey` knob and related hashing is removed.
Signed-off-by: eopXD <yuehtingc@nvidia.com>
Signed-off-by: Robin Kobus <19427718+Funatiq@users.noreply.github.com>
Signed-off-by: Daniel Campora <961215+dcampora@users.noreply.github.com>
Co-authored-by: Daniel Campora <961215+dcampora@users.noreply.github.com>
- Moved sorting related logic to a dedicated function for better clarity and maintainability.
- Enhanced sorting logic to separate finished context requests from ongoing ones before sorting by Lora task ID.
- Updated function documentation to reflect the sorting behavior and its purpose.
Signed-off-by: Robin Kobus <19427718+Funatiq@users.noreply.github.com>
- Only allocate additional outputs on last pipeline parallel rank in trtGptModelInflightBatching and executorImpl.
Signed-off-by: Robin Kobus <19427718+Funatiq@users.noreply.github.com>
* refactor: CreateNewDecoderRequests
Signed-off-by: Robin Kobus <19427718+Funatiq@users.noreply.github.com>
* refactor: Consolidate request generation in CreateNewDecoderRequests
- Removed the GenerateRequestOptions class and integrated its functionality into CreateNewDecoderRequests.
- Updated the constructor of CreateNewDecoderRequests to accept parameters for speculative decoding and normalization options.
- Modified the operator() method to handle request generation directly, improving code organization and reducing redundancy.
- Cleaned up associated includes and references throughout the codebase.
Signed-off-by: Robin Kobus <19427718+Funatiq@users.noreply.github.com>
* refactor: Simplify request handling in CreateNewDecoderRequests
- Removed the generateRequestOptions method and integrated its logic directly into the operator() method.
- Updated the request generation process to improve clarity and reduce redundancy.
- Adjusted the return type to streamline the handling of batch slots, decoder requests, and sampling configurations.
Signed-off-by: Robin Kobus <19427718+Funatiq@users.noreply.github.com>
* refactor: Enhance createDecoderRequests method in CreateNewDecoderRequests
- Updated the createDecoderRequests method to include additional parameters for decoder state and CUDA streams, improving flexibility in request handling.
- Removed redundant request generation logic from the operator() method, streamlining the process.
- Adjusted the newRequest method to utilize the updated decoder request structure, enhancing clarity and maintainability.
Signed-off-by: Robin Kobus <19427718+Funatiq@users.noreply.github.com>
* refactor: Use MedusaBuffers instead of RuntimeBuffers in CreateNewDecoderRequests
- Updated references from RuntimeBuffers to MedusaBuffers across the CreateNewDecoderRequests class and its methods, enhancing clarity in buffer management.
- Adjusted method signatures and internal logic to accommodate the new MedusaBuffers type, ensuring compatibility with existing functionality.
- Cleaned up unnecessary includes and improved code organization for better maintainability.
Signed-off-by: Robin Kobus <19427718+Funatiq@users.noreply.github.com>
* refactor: Update CreateNewDecoderRequests to use DecoderState and CudaStream parameters
- Modified method signatures in CreateNewDecoderRequests to replace GptDecoderBatched with runtime::decoder::DecoderState and added a separate CudaStream for the decoder.
- Adjusted the implementation of the operator() method to accommodate the new parameters, enhancing flexibility in request handling.
- Updated associated bindings in the pybind11 interface to reflect the changes in method signatures, ensuring consistency across the codebase.
Signed-off-by: Robin Kobus <19427718+Funatiq@users.noreply.github.com>
* refactor: Update TRTLLMSampler to use refactored create_new_decoder_requests
- Updated the sampler.py to reflect changes in the request handling logic, replacing generate_request_options with create_new_decoder_requests for improved clarity and consistency.
- Updated bindings and method signatures for decoder stream handling.
Signed-off-by: Robin Kobus <19427718+Funatiq@users.noreply.github.com>
* refactor: Update gptDecoderBatchedTest to use CreateNewDecoderRequests::newRequest
Signed-off-by: Robin Kobus <19427718+Funatiq@users.noreply.github.com>
---------
Signed-off-by: Robin Kobus <19427718+Funatiq@users.noreply.github.com>
