⚡ Bolt: cache inv_freq in rotary embeddings #6472
Open
+125
−88
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This change optimizes the calculation of `inv_freq` in rotary embedding layers by caching the result using `functools.lru_cache`. This avoids redundant tensor creation and mathematical operations (pow, division) during every forward pass or batch initialization, particularly benefiting multimodal models where these layers are frequently instantiated or called. - Introduce `get_inv_freq` cached function in `fastdeploy/model_executor/layers/rotary_embedding.py`. - Update `ErnieRotaryEmbedding`, `GlmRotaryEmbedding`, `QwenRotaryEmbedding` to use cached `inv_freq`. - Update `ErnieVlRotaryEmbedding3D` and `QwenVlRotaryEmbedding3D` to use cached `inv_freq`. Co-authored-by: ZeyuChen <1371212+ZeyuChen@users.noreply.github.com>
Contributor
|
👋 Jules, reporting for duty! I'm here to lend a hand with this pull request. When you start a review, I'll add a 👀 emoji to each comment to let you know I've read it. I'll focus on feedback directed at me and will do my best to stay out of conversations between you and other bots or reviewers to keep the noise down. I'll push a commit with your requested changes shortly after. Please note there might be a delay between these steps, but rest assured I'm on the job! For more direct control, you can switch me to Reactive Mode. When this mode is on, I will only act on comments where you specifically mention me with New to Jules? Learn more at jules.google/docs. For security, I will only act on instructions from the user who triggered this task. |
|
|
|
Thanks for your contribution! |
This change optimizes the calculation of `inv_freq` in rotary embedding layers by caching the result using `functools.lru_cache`. This avoids redundant tensor creation and mathematical operations (pow, division) during every forward pass or batch initialization, particularly benefiting multimodal models where these layers are frequently instantiated or called. - Introduce `get_inv_freq` cached function in `fastdeploy/model_executor/layers/rotary_embedding.py`. - Update `ErnieRotaryEmbedding`, `GlmRotaryEmbedding`, `QwenRotaryEmbedding` to use cached `inv_freq`. - Update `ErnieVlRotaryEmbedding3D` and `QwenVlRotaryEmbedding3D` to use cached `inv_freq`. Co-authored-by: ZeyuChen <1371212+ZeyuChen@users.noreply.github.com>
This change optimizes the calculation of `inv_freq` in rotary embedding layers by caching the result using `functools.lru_cache`. This avoids redundant tensor creation and mathematical operations (pow, division) during every forward pass or batch initialization, particularly benefiting multimodal models where these layers are frequently instantiated or called. - Introduce `get_inv_freq` cached function in `fastdeploy/model_executor/layers/rotary_embedding.py`. - Update `ErnieRotaryEmbedding`, `GlmRotaryEmbedding`, `QwenRotaryEmbedding` to use cached `inv_freq`. - Update `ErnieVlRotaryEmbedding3D` and `QwenVlRotaryEmbedding3D` to use cached `inv_freq`. Co-authored-by: ZeyuChen <1371212+ZeyuChen@users.noreply.github.com>
This change optimizes the calculation of `inv_freq` in rotary embedding layers by caching the result using `functools.lru_cache`. This avoids redundant tensor creation and mathematical operations (pow, division) during every forward pass or batch initialization, particularly benefiting multimodal models where these layers are frequently instantiated or called. - Introduce `get_inv_freq` cached function in `fastdeploy/model_executor/layers/rotary_embedding.py`. - Update `ErnieRotaryEmbedding`, `GlmRotaryEmbedding`, `QwenRotaryEmbedding` to use cached `inv_freq`. - Update `ErnieVlRotaryEmbedding3D` and `QwenVlRotaryEmbedding3D` to use cached `inv_freq`. Co-authored-by: ZeyuChen <1371212+ZeyuChen@users.noreply.github.com>
This change optimizes the calculation of `inv_freq` in rotary embedding layers by caching the result using `functools.lru_cache`. This avoids redundant tensor creation and mathematical operations (pow, division) during every forward pass or batch initialization, particularly benefiting multimodal models where these layers are frequently instantiated or called. - Introduce `get_inv_freq` cached function in `fastdeploy/model_executor/layers/rotary_embedding.py`. - Update `ErnieRotaryEmbedding`, `GlmRotaryEmbedding`, `QwenRotaryEmbedding` to use cached `inv_freq`. - Update `ErnieVlRotaryEmbedding3D` and `QwenVlRotaryEmbedding3D` to use cached `inv_freq`. - Fix: Ensure cached tensors are created on the correct device by using `position_ids.place` as cache key. Co-authored-by: ZeyuChen <1371212+ZeyuChen@users.noreply.github.com>
This change optimizes the calculation of `inv_freq` in rotary embedding layers by caching the result using `functools.lru_cache`. This avoids redundant tensor creation and mathematical operations (pow, division) during every forward pass or batch initialization, particularly benefiting multimodal models where these layers are frequently instantiated or called. Additionally, this change fixes the HPU CI failure by guarding the access to `paddle.compat.enable_torch_proxy` in `fastdeploy/__init__.py`. - Introduce `get_inv_freq` cached function in `fastdeploy/model_executor/layers/rotary_embedding.py`. - Update `ErnieRotaryEmbedding`, `GlmRotaryEmbedding`, `QwenRotaryEmbedding` to use cached `inv_freq`. - Update `ErnieVlRotaryEmbedding3D` and `QwenVlRotaryEmbedding3D` to use cached `inv_freq`. - Fix: Ensure cached tensors are created on the correct device by using `position_ids.place` as cache key. - Fix: Add guard for `paddle.compat` in `fastdeploy/__init__.py`. Co-authored-by: ZeyuChen <1371212+ZeyuChen@users.noreply.github.com>
This change optimizes the calculation of `inv_freq` in rotary embedding layers by caching the result using `functools.lru_cache`. This avoids redundant tensor creation and mathematical operations (pow, division) during every forward pass or batch initialization, particularly benefiting multimodal models where these layers are frequently instantiated or called. Additionally, this change fixes the HPU CI failure by guarding the access to `paddle.compat.enable_torch_proxy` in `fastdeploy/__init__.py`. - Introduce `get_inv_freq` cached function in `fastdeploy/model_executor/layers/rotary_embedding.py`. - Update `ErnieRotaryEmbedding`, `GlmRotaryEmbedding`, `QwenRotaryEmbedding` to use cached `inv_freq`. - Update `ErnieVlRotaryEmbedding3D` and `QwenVlRotaryEmbedding3D` to use cached `inv_freq`. - Fix: Ensure cached tensors are created on the correct device by using `position_ids.place` as cache key. - Fix: Add guard for `paddle.compat` in `fastdeploy/__init__.py`. Co-authored-by: ZeyuChen <1371212+ZeyuChen@users.noreply.github.com>
- Cached `inv_freq` calculation in `fastdeploy/model_executor/layers/rotary_embedding.py`. - Guarded `paddle.compat` usage to fix HPU CI failures. Co-authored-by: ZeyuChen <1371212+ZeyuChen@users.noreply.github.com>
- Cached `inv_freq` calculation in `fastdeploy/model_executor/layers/rotary_embedding.py` for standard implementations (`ErnieRotaryEmbedding`, `GlmRotaryEmbedding`, `QwenRotaryEmbedding`). - Guarded `paddle.compat` usage across the codebase to fix CI failures on environments missing `paddle.compat`. - Retained original logic for `ErnieVlRotaryEmbedding3D` and `QwenVlRotaryEmbedding3D` to ensure stability on MetaX/MACA. Co-authored-by: ZeyuChen <1371212+ZeyuChen@users.noreply.github.com>
- Cached `inv_freq` calculation in `fastdeploy/model_executor/layers/rotary_embedding.py` for standard implementations (`ErnieRotaryEmbedding`, `GlmRotaryEmbedding`, `QwenRotaryEmbedding`). - Guarded `paddle.compat` usage across the codebase to fix CI failures on environments missing `paddle.compat`. - Retained original logic for `ErnieVlRotaryEmbedding3D` and `QwenVlRotaryEmbedding3D` to ensure stability on MetaX/MACA. Co-authored-by: ZeyuChen <1371212+ZeyuChen@users.noreply.github.com>
- Cached `inv_freq` calculation in `fastdeploy/model_executor/layers/rotary_embedding.py` for `GlmRotaryEmbedding` and `QwenRotaryEmbedding`. - Reverted `ErnieRotaryEmbedding`, `ErnieVlRotaryEmbedding3D`, and `QwenVlRotaryEmbedding3D` to inline logic to ensure stability on Iluvatar/MetaX environments. - Guarded `paddle.compat` usage across the codebase to fix CI failures on environments missing `paddle.compat`. Co-authored-by: ZeyuChen <1371212+ZeyuChen@users.noreply.github.com>
- Cached `inv_freq` calculation in `fastdeploy/model_executor/layers/rotary_embedding.py` for `GlmRotaryEmbedding` and `QwenRotaryEmbedding`. - Reverted `ErnieRotaryEmbedding`, `ErnieVlRotaryEmbedding3D`, and `QwenVlRotaryEmbedding3D` to inline logic to ensure stability on Iluvatar/MetaX environments. - Guarded `paddle.compat` usage across the codebase to fix CI failures on environments missing `paddle.compat`. - Added fallback for `paddle.nn.functional.swiglu` in `fastdeploy/model_executor/ops/iluvatar/moe_ops.py` to fix Iluvatar CI import errors. Co-authored-by: ZeyuChen <1371212+ZeyuChen@users.noreply.github.com>
ZeyuChen
requested a deployment
to
Metax_ci
GitHub Actions
In progress
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
💡 What: Cached the calculation of
inv_freq(inverse frequency) tensors in Rotary Embedding layers usingfunctools.lru_cache.🎯 Why: The
inv_freqtensor depends only onrotary_dim,base, and the device. In the current implementation, it is recomputed every time__call__is invoked (for some layers) or every time a new layer instance is created (which happens frequently for multimodal models inGPUModelRunner._process_mm_features). Caching this computation saves tensor allocation and arithmetic operations.📊 Impact: Reduces overhead in the hot path of rotary embedding calculation. For multimodal models,
get_ropeis called for every batch, so this avoids O(N) operations per batch.🔬 Measurement: Verified by code inspection that redundant calculations are replaced by a cached function call. The cached function ensures the tensor is returned on the correct device.
PR created automatically by Jules for task 13163577974838169073 started by @ZeyuChen