[Iluvatar GPU] Adapt VL model #4313
Open
+344
−228
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
|
Thanks for your contribution! |
wuyujiji
force-pushed
the
adapt_vl
branch
6 times, most recently
from
b5fe3b3 to
d0a687a
Compare
There was a problem hiding this comment.
Pull Request Overview
This PR adapts VL (Vision-Language) models for Iluvatar GPU hardware, implementing platform-specific optimizations and configurations to support multimodal inference on Iluvatar devices.
- Pin paddleformers version to 0.3.0 for compatibility
- Implement Iluvatar-specific attention backend optimizations for VL models
- Add support for text-image processing operations and memory management
- Provide comprehensive documentation with installation and usage examples
Reviewed Changes
Copilot reviewed 12 out of 12 changed files in this pull request and generated 3 comments.
Show a summary per file
| File | Description |
|---|---|
| requirements_iluvatar.txt | Pin paddleformers to version 0.3.0 for stability |
| fastdeploy/worker/iluvatar_worker.py | Configure Paddle flags for multimodal support |
| fastdeploy/worker/iluvatar_model_runner.py | Add VL model initialization and rope embedding handling |
| fastdeploy/worker/gpu_model_runner.py | Import additional Iluvatar-specific operations |
| fastdeploy/model_executor/models/ernie4_5_vl/modeling_resampler.py | Disable fused matmul bias for Iluvatar platform |
| fastdeploy/model_executor/models/ernie4_5_vl/image_op.py | Add Iluvatar platform support for image operations |
| fastdeploy/model_executor/models/ernie4_5_vl/dfnrope/modeling.py | Disable fused matmul bias for Iluvatar platform |
| fastdeploy/model_executor/layers/rotary_embedding.py | Add Iluvatar-specific rotary embedding handling |
| fastdeploy/model_executor/layers/attention/iluvatar_attn_backend.py | Implement VL-specific attention metadata and tensor handling |
| docs/zh/get_started/installation/iluvatar_gpu.md | Add Chinese documentation for VL model usage |
| docs/get_started/installation/iluvatar_gpu.md | Add English documentation for VL model usage |
| custom_ops/setup_ops.py | Register additional CUDA operations for Iluvatar support |
YuanRisheng
previously approved these changes
Oct 15, 2025
wuyujiji
force-pushed
the
adapt_vl
branch
2 times, most recently
from
a58db43 to
f2c3ff5
Compare
YuanRisheng
previously approved these changes
Oct 16, 2025
yuanlehome
reviewed
Oct 16, 2025
yuanlehome
reviewed
Oct 16, 2025
custom_ops/setup_ops.py
Outdated
| "gpu_ops/sample_kernels/top_k_renorm_probs.cu", | ||
| "gpu_ops/text_image_index_out.cu", | ||
| "gpu_ops/text_image_gather_scatter.cu", | ||
| "gpu_ops/extract_text_token_output.cu", |
There was a problem hiding this comment.
extract_text_token_output已经被废弃，麻烦一并给删掉吧，包括算子实现、单测等等
There was a problem hiding this comment.
@yuanlehome 这里指的是删掉custom_ops/gpu_ops/extract_text_token_output.cu，test/operators/test_extract_text_token_output.py和cpp_extentions.cc里关于extract_text_token_output的注册？除了这三处之外，setup_ops.py里"gpu_ops/extract_text_token_output.cu"在metax_gpu里也用到了，我如果删了cu文件的实现会对这里造成影响吧
There was a problem hiding this comment.
@yuanlehome 已经删了，麻烦看下删的对不对
wuyujiji
dismissed stale reviews from gongshaotian and YuanRisheng
via
002cbd8
heavengate
reviewed
Oct 16, 2025
| @@ -1,101 +0,0 @@ | |||
| // Copyright (c) 2024 PaddlePaddle Authors. All Rights Reserved. | |||
yuanlehome
approved these changes
Oct 16, 2025
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
5 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.
在天数硬件上适配VL模型