[MM][Perf][CG] Support dual-path ViT full CUDA graph for DeepSeek-OCR (#43586)
Signed-off-by: shen-shanshan <467638484@qq.com> Signed-off-by: Isotr0py <mozf@mail2.sysu.edu.cn> Co-authored-by: Roger Wang <hey@rogerw.io> Co-authored-by: Isotr0py <mozf@mail2.sysu.edu.cn>
This commit is contained in:
co-authored by
Roger Wang
Isotr0py
parent
b2cfae777d
commit
ad32608e24
@@ -2,6 +2,8 @@
|
||||
|
||||
The [CUDA Graphs](cuda_graphs.md) infrastructure in vLLM primarily targets the **decoder** (language model) forward pass. vLLM also supports capturing the **encoder** (vision transformer) forward pass as CUDA Graphs, independently from the decoder. This is based on <https://github.com/vllm-project/vllm/pull/35963>.
|
||||
|
||||
For two-tower vision encoders (e.g., DeepSeek-OCR's SAM + CLIP with dynamic tiling), a **dual-path graph** mode captures two independent sets of CUDA graphs — one for the global image path and one for the local patch path — enabling independent budget selection and partial eager fallback per path. This is based on <https://github.com/vllm-project/vllm/pull/43586>.
|
||||
|
||||
!!! note
|
||||
Encoder CUDA Graphs are orthogonal to decoder CUDA Graphs — both can be enabled simultaneously. Encoder graphs capture the vision encoder execution (e.g., ViT in Qwen3-VL), while decoder graphs capture the language model execution as described in the [CUDA Graphs design document](cuda_graphs.md).
|
||||
|
||||
@@ -11,6 +13,8 @@ Vision encoder inference incurs CUDA kernel launch overhead on the host side. Th
|
||||
|
||||
Encoder CUDA Graphs eliminate this overhead by pre-capturing the full encoder forward pass at multiple token budget levels during model initialization, then replaying the appropriate graph at runtime.
|
||||
|
||||
For two-tower vision encoders such as DeepSeek-OCR (SAM + CLIP with dynamic tiling), the global image path and local patch path have independent token profiles (272 tokens per global image vs. 100 tokens per local patch). Capturing a single monolithic graph for both paths would significantly reduce packing efficiency. The dual-path graph mode captures each path as a separate set of budgets, allowing the manager to pack and replay each path independently.
|
||||
|
||||
## Design
|
||||
|
||||
The encoder CUDA Graph system uses a **budget-based capture/replay** strategy, managed by [EncoderCudaGraphManager][vllm.v1.worker.encoder_cudagraph.EncoderCudaGraphManager]. The system contains the following core components:
|
||||
@@ -37,10 +41,14 @@ class BudgetGraphMetadata:
|
||||
|
||||
Budgets are auto-generated as power-of-2 levels from a model-provided range via `get_encoder_cudagraph_budget_range()`, with the maximum budget always included even if it does not fall on a power-of-2 boundary. Budgets can also be explicitly specified by the user via `encoder_cudagraph_token_budgets` in `CompilationConfig`.
|
||||
|
||||
When `EncoderCudaGraphConfig.enable_dual_path_graph` is `True`, the manager generates two independent budget lists — `global_token_budgets` (multiples of `global_token_per_image`) and `local_token_budgets` (multiples of `local_token_per_patch`) — and stores captured graphs under `budget_graphs["global"]` and `budget_graphs["local"]` respectively.
|
||||
|
||||
### Greedy bin-packing at runtime
|
||||
|
||||
When a batch of images arrives, the manager sorts images by output token count (smallest first) and greedily packs as many images as possible into each sub-batch while staying within the **largest** token budget and the maximum batch size. Once a sub-batch is finalized (the next image would overflow either constraint), the manager finds the **smallest** budget that fits the sub-batch's total tokens and replays the corresponding CUDA Graph. This repeats until the batch is exhausted. Images that exceed all budgets fall back to eager execution.
|
||||
|
||||
For dual-path models, the manager routes to `_execute_local_dual_path()`, which constrains both global and local token budgets simultaneously during packing (see [Dual-Path graph capture](#dual-path-graph-capture)).
|
||||
|
||||
For each graph replay:
|
||||
|
||||
1. Call `prepare_encoder_cudagraph_replay_buffers()` to compute buffer values (including `pixel_values` and precomputed metadata) from actual batch inputs.
|
||||
@@ -48,6 +56,42 @@ For each graph replay:
|
||||
3. Replay the CUDA Graph.
|
||||
4. Clone outputs from `output_buffer` (cloning is necessary since the buffer is reused across replays).
|
||||
|
||||
### Dual-Path graph capture
|
||||
|
||||
For two-tower vision encoders (e.g., DeepSeek-OCR), the `EncoderCudaGraphConfig` sets `enable_dual_path_graph=True` and provides `global_token_per_image` / `local_token_per_patch`. The manager captures two independent sets of CUDA graphs — one for the **global** image path and one for the **local** patch path — stored under `budget_graphs["global"]` and `budget_graphs["local"]` respectively.
|
||||
|
||||
**Budget generation.** Two separate budget lists are generated:
|
||||
|
||||
* `global_token_budgets` — power-of-2 multiples of `global_token_per_image` (e.g., `[272, 544, 1088, 2176, 4352, 8704, 13824]` for DeepSeek-OCR).
|
||||
* `local_token_budgets` — power-of-2 multiples of `local_token_per_patch` (e.g., `[0, 100, 200, 400, 800, 1600, 3200, 6400, 12800]` for DeepSeek-OCR). A budget of `0` is always included to handle images with no local patches (images ≤ 640×640 that produce only global features).
|
||||
|
||||
Both lists are capped at the same `max_budget`.
|
||||
|
||||
**Dual-path greedy packing.** Each `EncoderItemSpec` provides both `global_output_tokens` (constant per image) and `local_output_tokens` (proportional to the patch count). The dual-path packing algorithm constrains both budgets simultaneously:
|
||||
|
||||
* Sort images by total output tokens (global + local), smallest first.
|
||||
* Greedily pack images: an image is added to the current sub-batch only if both the accumulated global tokens ≤ `max_global_budget` **and** the accumulated local tokens ≤ `max_local_budget`, with the image count ≤ `max_batch_size`.
|
||||
* Once either constraint would overflow, finalize the sub-batch and find the smallest fitting budget **independently** for each path.
|
||||
* Repeat until all images are packed.
|
||||
|
||||
**Partial graph fallback.** After packing, each sub-batch falls into one of four execution scenarios:
|
||||
|
||||
| Global budget | Local budget | Execution |
|
||||
| :---: | :---: | --- |
|
||||
| Found | Found | Both paths use CUDA graph replay |
|
||||
| Found | `None` | Global graph replay + local path skipped (no patches) |
|
||||
| `None` | Found | Global eager fallback + local graph replay |
|
||||
| `None` | `None` | Both paths fall back to eager execution |
|
||||
|
||||
Note that the `0`-budget graph is never actually replayed for local — it signals that local patch processing should be skipped entirely.
|
||||
|
||||
**Buffer keys per path.** Global and local paths use different buffer keys. For DeepSeek-OCR, the global path uses `pixel_values` (full images, shape `[B, 3, 1280, 1280]`) while the local path uses `images_crop` (patches, shape `[P, 3, 1024, 1024]`). The manager iterates over each captured graph's own `input_buffers.keys()` rather than a shared `buffer_keys` list, so both paths can use different buffers.
|
||||
|
||||
**Post-processing.** The `postprocess_encoder_output` method receives a `local_output` parameter (a tensor or `None`) containing the local-path encoder output. The model is responsible for assembling global and local features into the final per-image embedding. For DeepSeek-OCR, this means reshaping the global output into `[B, 272, n_embed]`, the local output into `[P, 100, n_embed]`, assembling patch grids with newline tokens, and concatenating `[patches_grid, global, view_separator]` for each image.
|
||||
|
||||
!!! note
|
||||
The dual-path design enables partial CUDA graph coverage — one path can hit while the other falls back to eager. This avoids wasted compute on zero-padded patch buffers for untiled images and avoids graph invalidation caused by variable `crop_shape` per image.
|
||||
|
||||
### Data-parallel support
|
||||
|
||||
When `mm_encoder_tp_mode="data"`, the manager distributes images across TP ranks using load-balanced assignment via `get_load_balance_assignment`, executes locally on each rank, then gathers results back in the original order via `tensor_model_parallel_all_gather`.
|
||||
@@ -67,29 +111,30 @@ Models opt-in to encoder CUDA Graphs by implementing the [SupportsEncoderCudaGra
|
||||
|
||||
* `get_encoder_cudagraph_config()` — returns static configuration (supported modalities, buffer keys, output hidden size, padding logics, max frames per video).
|
||||
* `get_encoder_cudagraph_budget_range(vllm_config)` — returns `(min_budget, max_budget)` for auto-inference of token budgets.
|
||||
* `get_encoder_cudagraph_item_specs(mm_kwargs)` — returns `list[EncoderItemSpec]` describing each item with its input size and output token count. Replaces the former three separate methods (`get_num_items`, `get_per_item_output_tokens`, `get_per_item_input_sizes`).
|
||||
* `get_encoder_cudagraph_item_specs(mm_kwargs)` — returns `list[EncoderItemSpec]` describing each item with its input size, total output token count (`output_tokens`), and optionally per-path token counts (`global_output_tokens`, `local_output_tokens`) for dual-path models.
|
||||
* `select_encoder_cudagraph_items(mm_kwargs, indices)` — extracts a sub-batch of items by index, used during greedy packing and DP sharding.
|
||||
* `prepare_encoder_cudagraph_capture_inputs(...)` — creates dummy inputs for graph capture. Returns `EncoderCudaGraphCaptureInputs` with a single `values: dict[str, torch.Tensor]` that contains all buffers to be recorded into the graph.
|
||||
* `prepare_encoder_cudagraph_replay_buffers(mm_kwargs, max_batch_size, max_frames_per_batch)` — computes buffer values from actual batch inputs. Returns `EncoderCudaGraphReplayBuffers` with a `values` dict whose keys match `buffer_keys` in the config.
|
||||
* `encoder_cudagraph_forward(inputs: dict[str, torch.Tensor])` — forward pass accepting only fixed-shaped input tensors (the captured `values` dict). Called during both capture and replay. The `pixel_values` tensor is included in `inputs` alongside metadata buffers.
|
||||
* `encoder_eager_forward(mm_kwargs)` — fallback eager forward when no graph fits.
|
||||
* `postprocess_encoder_output(...)` — post-process encoder output, delegates to `scatter_output_slices` by default.
|
||||
* `prepare_encoder_cudagraph_capture_inputs(..., path="default")` — creates dummy inputs for graph capture. The `path` parameter (`"global"` or `"local"`) tells the model which path to generate dummy inputs for. Returns `EncoderCudaGraphCaptureInputs` with a single `values: dict[str, torch.Tensor]` that contains all buffers to be recorded into the graph.
|
||||
* `prepare_encoder_cudagraph_replay_buffers(mm_kwargs, max_batch_size, max_frames_per_batch, path="default")` — computes buffer values from actual batch inputs. The `path` parameter selects which modality keys to extract from `mm_kwargs`. Returns `EncoderCudaGraphReplayBuffers` with a `values` dict whose keys match the captured graph's `input_buffers.keys()`.
|
||||
* `encoder_cudagraph_forward(inputs: dict[str, torch.Tensor], path="default")` — forward pass accepting only fixed-shaped input tensors (the captured `values` dict). Called during both capture and replay. The `path` parameter dispatches to the correct encoder sub-module (e.g., global vs. local path for DeepSeek-OCR).
|
||||
* `encoder_eager_forward(mm_kwargs, path="default")` — fallback eager forward when no graph fits. When `path` is `"global"` or `"local"`, runs only that encoder path without graph capture.
|
||||
* `postprocess_encoder_output(..., local_output=None)` — post-process encoder output. The `local_output` parameter receives the local-path encoder output tensor (or `None`), enabling dual-path models to assemble global and local features into the final per-image embedding.
|
||||
|
||||
!!! note
|
||||
The `SupportsEncoderCudaGraph` protocol is designed to be model-agnostic. New vision encoder models can opt-in by implementing the protocol methods without modifying the manager.
|
||||
|
||||
**Supported models:**
|
||||
|
||||
| Architecture | Models | CG for Image | CG for Video |
|
||||
| ------------ | ------ | ------------ | ------------ |
|
||||
| `Llama4ForConditionalGeneration` | `Llama 4` | ✅︎ | - |
|
||||
| `InternVLChatModel` | `InternVL3.5`, `InternVL3`, `InternVL2.5`, `InternVL2` | ✅︎ | ✅︎ |
|
||||
| `Qwen2VLForConditionalGeneration` | `Qwen2-VL` | ✅︎ | ✅︎ |
|
||||
| `Qwen2_5_VLForConditionalGeneration` | `Qwen2.5-VL` | ✅︎ | ✅︎ |
|
||||
| `Qwen3VLForConditionalGeneration` | `Qwen3-VL` | ✅︎ | ✅︎ |
|
||||
| `Qwen3_5ForConditionalGeneration` | `Qwen3.5` | ✅︎ | ✅︎ |
|
||||
| `Step3VLForConditionalGeneration` | `Step3-VL` | ✅︎ | ❌︎ |
|
||||
| `Glm4vForConditionalGeneration` | `GLM-4.1V, GLM-4.6V-Flash` | ✅︎ | ✅︎ |
|
||||
| Architecture | Models | CG for Image | CG for Video | Dual-Path Graph |
|
||||
| ------------ | ------ | ------------ | ------------ | --------------- |
|
||||
| `DeepseekOCRForCausalLM` | `DeepSeek-OCR` | ✅︎ | ❌︎ | ✅︎ |
|
||||
| `Glm4vForConditionalGeneration` | `GLM-4.1V, GLM-4.6V-Flash` | ✅︎ | ✅︎ | ❌︎ |
|
||||
| `InternVLChatModel` | `InternVL3.5`, `InternVL3`, `InternVL2.5`, `InternVL2` | ✅︎ | ✅︎ | ❌︎ |
|
||||
| `Llama4ForConditionalGeneration` | `Llama 4` | ✅︎ | ❌︎ | ❌︎ |
|
||||
| `Qwen2VLForConditionalGeneration` | `Qwen2-VL` | ✅︎ | ✅︎ | ❌︎ |
|
||||
| `Qwen2_5_VLForConditionalGeneration` | `Qwen2.5-VL` | ✅︎ | ✅︎ | ❌︎ |
|
||||
| `Qwen3VLForConditionalGeneration` | `Qwen3-VL` | ✅︎ | ✅︎ | ❌︎ |
|
||||
| `Qwen3_5ForConditionalGeneration` | `Qwen3.5` | ✅︎ | ✅︎ | ❌︎ |
|
||||
| `Step3VLForConditionalGeneration` | `Step3-VL` | ✅︎ | ❌︎ | ❌︎ |
|
||||
|
||||
!!! note
|
||||
Encoder CUDA Graphs have currently been tested with `--mm-encoder-attn-backend=FLASH_ATTN` and `--mm-encoder-attn-backend=FLASHINFER` on Blackwell GPUs.
|
||||
@@ -104,6 +149,8 @@ Three fields in `CompilationConfig` control encoder CUDA Graphs:
|
||||
* `encoder_cudagraph_max_vision_items_per_batch` (`int`, default `0`) — maximum number of images/videos per batch during capture. If 0 (default), auto-inferred as `max_budget // min_budget`.
|
||||
* `encoder_cudagraph_max_frames_per_batch` (`int`, default `None`) — maximum number of video frames per batch during capture. If `None` (default), auto-inferred as `encoder_cudagraph_max_vision_items_per_batch * max_frames_per_video` (`max_frames_per_video` is a model-specific value from `EncoderCudaGraphConfig`, computed by `get_max_frames_per_video()` on the model). If we limit the video count per prompt to `0`, it will also be set to `0` (i.e., fall back to image-only mode).
|
||||
|
||||
Dual-path mode is configured at the model level via `EncoderCudaGraphConfig` fields (`enable_dual_path_graph`, `global_token_per_image`, `local_token_per_patch`) — no additional user configuration is required. The manager automatically generates separate budget lists and routes to dual-path execution when the model opts in.
|
||||
|
||||
## Usage guide
|
||||
|
||||
### Image inference
|
||||
|
||||
Reference in New Issue
Block a user