Implement 2d tiled matmulnbits specialized for prefill #23058
Merged
Conversation
This file contains 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
|
/azp run ONNX Runtime Web CI Pipeline,Windows GPU CI Pipeline,Linux Android Emulator QNN CI Pipeline |
|
/azp run Linux CPU CI Pipeline,Linux CPU Minimal Build E2E CI Pipeline,Linux GPU CI Pipeline,Linux GPU TensorRT CI Pipeline,Linux OpenVINO CI Pipeline,Linux QNN CI Pipeline,MacOS CI Pipeline,Windows ARM64 QNN CI Pipeline,Windows CPU CI Pipeline |
|
Azure Pipelines successfully started running 2 pipeline(s). |
|
/azp run Windows GPU TensorRT CI Pipeline,onnxruntime-binary-size-checks-ci-pipeline,orttraining-linux-ci-pipeline,orttraining-linux-gpu-ci-pipeline,orttraining-ortmodule-distributed,Windows x64 QNN CI Pipeline,Big Models |
|
/azp run Windows GPU CUDA CI Pipeline,Windows GPU DML CI Pipeline,Windows GPU Doc Gen CI Pipeline |
|
Azure Pipelines successfully started running 4 pipeline(s). |
|
Azure Pipelines successfully started running 3 pipeline(s). |
|
Azure Pipelines successfully started running 9 pipeline(s). |
guschmue
previously approved these changes
Dec 10, 2024
|
just needs a fix for the macos build |
|
/azp run ONNX Runtime Web CI Pipeline,Windows GPU CI Pipeline,Linux Android Emulator QNN CI Pipeline |
|
Azure Pipelines successfully started running 2 pipeline(s). |
|
/azp run Linux CPU CI Pipeline,Linux CPU Minimal Build E2E CI Pipeline,Linux GPU CI Pipeline,Linux GPU TensorRT CI Pipeline,Linux OpenVINO CI Pipeline,Linux QNN CI Pipeline,MacOS CI Pipeline,Windows ARM64 QNN CI Pipeline,Windows CPU CI Pipeline |
|
/azp run Windows GPU TensorRT CI Pipeline,onnxruntime-binary-size-checks-ci-pipeline,orttraining-linux-ci-pipeline,orttraining-linux-gpu-ci-pipeline,orttraining-ortmodule-distributed,Windows x64 QNN CI Pipeline,Big Models |
|
/azp run Windows GPU CUDA CI Pipeline,Windows GPU DML CI Pipeline,Windows GPU Doc Gen CI Pipeline |
|
Azure Pipelines successfully started running 4 pipeline(s). |
|
Azure Pipelines successfully started running 3 pipeline(s). |
|
Azure Pipelines successfully started running 9 pipeline(s). |
|
I see similar prefill improvements on Xe TGL, ~1.5x |
guschmue
approved these changes
Dec 11, 2024
ankitm3k
pushed a commit
to intel/onnxruntime
that referenced
this pull request
Dec 11, 2024
### Description
This change implements matmul4bits with tiling both for A and B. This is
beneficial for prefill scenarios on Intel integrated GPUs, because each
row of A has to run through the same set of shared rows of B. This
change should improve core occupancy and model_benchmark does indicate
improvements for prefill.
The same shader is not used for generation because when A has just a
single row, the other threads in the workgroup get unused and that hurts
performance.
```
-- Baseline run on an Alderlake GPU --
C:\onnxruntime>C:\model_benchmark\model_benchmark.exe -i C:\Phi-3.5-mini-instruct-onnx-web\Phi-3.5-mini-instruct-onnx-web -l 500
Batch size: 1, prompt tokens: 501, tokens to generate: 128
Prompt processing (time to first token):
avg (us): 1.72338e+07
avg (tokens/s): 29.0707 <<
p50 (us): 1.72548e+07
stddev (us): 57012.8
n: 5 * 501 token(s)
Token generation:
avg (us): 79227.5
avg (tokens/s): 12.6219
p50 (us): 79284.4
stddev (us): 2109.72
n: 635 * 1 token(s)
Token sampling:
avg (us): 15.8198
avg (tokens/s): 63211.8
p50 (us): 14.3
stddev (us): 8.67178
n: 640 * 1 token(s)
E2E generation (entire generation loop):
avg (ms): 27297.8
p50 (ms): 27269.8
stddev (ms): 89.4322
n: 5
Peak working set size (bytes): 5490987008
WebGPU device lost (2): Device was destroyed.
----------------------------------- With Prefill Optimization ----
C:\onnxruntime>C:\model_benchmark\model_benchmark.exe -i C:\Phi-3.5-mini-instruct-onnx-web\Phi-3.5-mini-instruct-onnx-web -l 500
Batch size: 1, prompt tokens: 501, tokens to generate: 128
Prompt processing (time to first token):
avg (us): 1.2135e+07
avg (tokens/s): 41.2856 <<
p50 (us): 1.21288e+07
stddev (us): 21282.1
n: 5 * 501 token(s)
Token generation:
avg (us): 78945.3
avg (tokens/s): 12.667
p50 (us): 78900.7
stddev (us): 2232.43
n: 635 * 1 token(s)
Token sampling:
avg (us): 20.5608
avg (tokens/s): 48636.3
p50 (us): 18.7
stddev (us): 19.0409
n: 640 * 1 token(s)
E2E generation (entire generation loop):
avg (ms): 22163.8
p50 (ms): 22160.1
stddev (ms): 31.3122
n: 5
Peak working set size (bytes): 5478862848
WebGPU device lost (2): Device was destroyed.
```
ankitm3k
pushed a commit
to intel/onnxruntime
that referenced
this pull request
Dec 11, 2024
### Description
This change implements matmul4bits with tiling both for A and B. This is
beneficial for prefill scenarios on Intel integrated GPUs, because each
row of A has to run through the same set of shared rows of B. This
change should improve core occupancy and model_benchmark does indicate
improvements for prefill.
The same shader is not used for generation because when A has just a
single row, the other threads in the workgroup get unused and that hurts
performance.
```
-- Baseline run on an Alderlake GPU --
C:\onnxruntime>C:\model_benchmark\model_benchmark.exe -i C:\Phi-3.5-mini-instruct-onnx-web\Phi-3.5-mini-instruct-onnx-web -l 500
Batch size: 1, prompt tokens: 501, tokens to generate: 128
Prompt processing (time to first token):
avg (us): 1.72338e+07
avg (tokens/s): 29.0707 <<
p50 (us): 1.72548e+07
stddev (us): 57012.8
n: 5 * 501 token(s)
Token generation:
avg (us): 79227.5
avg (tokens/s): 12.6219
p50 (us): 79284.4
stddev (us): 2109.72
n: 635 * 1 token(s)
Token sampling:
avg (us): 15.8198
avg (tokens/s): 63211.8
p50 (us): 14.3
stddev (us): 8.67178
n: 640 * 1 token(s)
E2E generation (entire generation loop):
avg (ms): 27297.8
p50 (ms): 27269.8
stddev (ms): 89.4322
n: 5
Peak working set size (bytes): 5490987008
WebGPU device lost (2): Device was destroyed.
----------------------------------- With Prefill Optimization ----
C:\onnxruntime>C:\model_benchmark\model_benchmark.exe -i C:\Phi-3.5-mini-instruct-onnx-web\Phi-3.5-mini-instruct-onnx-web -l 500
Batch size: 1, prompt tokens: 501, tokens to generate: 128
Prompt processing (time to first token):
avg (us): 1.2135e+07
avg (tokens/s): 41.2856 <<
p50 (us): 1.21288e+07
stddev (us): 21282.1
n: 5 * 501 token(s)
Token generation:
avg (us): 78945.3
avg (tokens/s): 12.667
p50 (us): 78900.7
stddev (us): 2232.43
n: 635 * 1 token(s)
Token sampling:
avg (us): 20.5608
avg (tokens/s): 48636.3
p50 (us): 18.7
stddev (us): 19.0409
n: 640 * 1 token(s)
E2E generation (entire generation loop):
avg (ms): 22163.8
p50 (ms): 22160.1
stddev (ms): 31.3122
n: 5
Peak working set size (bytes): 5478862848
WebGPU device lost (2): Device was destroyed.
```
ankitm3k
pushed a commit
to intel/onnxruntime
that referenced
this pull request
Dec 11, 2024
### Description
This change implements matmul4bits with tiling both for A and B. This is
beneficial for prefill scenarios on Intel integrated GPUs, because each
row of A has to run through the same set of shared rows of B. This
change should improve core occupancy and model_benchmark does indicate
improvements for prefill.
The same shader is not used for generation because when A has just a
single row, the other threads in the workgroup get unused and that hurts
performance.
```
-- Baseline run on an Alderlake GPU --
C:\onnxruntime>C:\model_benchmark\model_benchmark.exe -i C:\Phi-3.5-mini-instruct-onnx-web\Phi-3.5-mini-instruct-onnx-web -l 500
Batch size: 1, prompt tokens: 501, tokens to generate: 128
Prompt processing (time to first token):
avg (us): 1.72338e+07
avg (tokens/s): 29.0707 <<
p50 (us): 1.72548e+07
stddev (us): 57012.8
n: 5 * 501 token(s)
Token generation:
avg (us): 79227.5
avg (tokens/s): 12.6219
p50 (us): 79284.4
stddev (us): 2109.72
n: 635 * 1 token(s)
Token sampling:
avg (us): 15.8198
avg (tokens/s): 63211.8
p50 (us): 14.3
stddev (us): 8.67178
n: 640 * 1 token(s)
E2E generation (entire generation loop):
avg (ms): 27297.8
p50 (ms): 27269.8
stddev (ms): 89.4322
n: 5
Peak working set size (bytes): 5490987008
WebGPU device lost (2): Device was destroyed.
----------------------------------- With Prefill Optimization ----
C:\onnxruntime>C:\model_benchmark\model_benchmark.exe -i C:\Phi-3.5-mini-instruct-onnx-web\Phi-3.5-mini-instruct-onnx-web -l 500
Batch size: 1, prompt tokens: 501, tokens to generate: 128
Prompt processing (time to first token):
avg (us): 1.2135e+07
avg (tokens/s): 41.2856 <<
p50 (us): 1.21288e+07
stddev (us): 21282.1
n: 5 * 501 token(s)
Token generation:
avg (us): 78945.3
avg (tokens/s): 12.667
p50 (us): 78900.7
stddev (us): 2232.43
n: 635 * 1 token(s)
Token sampling:
avg (us): 20.5608
avg (tokens/s): 48636.3
p50 (us): 18.7
stddev (us): 19.0409
n: 640 * 1 token(s)
E2E generation (entire generation loop):
avg (ms): 22163.8
p50 (ms): 22160.1
stddev (ms): 31.3122
n: 5
Peak working set size (bytes): 5478862848
WebGPU device lost (2): Device was destroyed.
```
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
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.
Description
This change implements matmul4bits with tiling both for A and B. This is beneficial for prefill scenarios on Intel integrated GPUs, because each row of A has to run through the same set of shared rows of B. This change should improve core occupancy and model_benchmark does indicate improvements for prefill.
The same shader is not used for generation because when A has just a single row, the other threads in the workgroup get unused and that hurts performance.