Port oss f16_fast_gemv into fbcode #3610
Open
+747
−2
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
|
This pull request was exported from Phabricator. Differential Revision: D68470488 |
✅ Deploy Preview for pytorch-fbgemm-docs ready!
To edit notification comments on pull requests, go to your Netlify site configuration. |
YUNQIUGUO
added a commit
to YUNQIUGUO/FBGEMM
that referenced
this pull request
Jan 24, 2025
Summary: X-link: facebookresearch/FBGEMM#688 This diff content includes: 1. Port OSS FastGEMV `fp16` kernel into fbcode and expose to python as a step 1 - `torch.ops.fbgemm.f16_fast_gemv` https://github.com/wangsiping97/FastGEMV/blob/1fdff6f74aade033c02727a419afd6a4b4bfbc3f/fast_gemv.cu#L14 2. Add `fp16_oss_fast_gemv` to quantize ops benchmark script 3. Add two simple tests for custom op`torch.ops.fbgemm.f16_fast_gemv` to test - `torch.compile()` able - correctness **Next step:** Need fp8 mixed precision support for fast gemv kernel which is what we want Differential Revision: D68470488
YUNQIUGUO
force-pushed
the
export-D68470488
branch
from
7e680f3 to
b031f69
Compare
|
This pull request was exported from Phabricator. Differential Revision: D68470488 |
YUNQIUGUO
added a commit
to YUNQIUGUO/FBGEMM
that referenced
this pull request
Jan 24, 2025
Summary: X-link: facebookresearch/FBGEMM#688 This diff content includes: 1. Port OSS FastGEMV `fp16` kernel into fbcode and expose to python as a step 1 - `torch.ops.fbgemm.f16_fast_gemv` https://github.com/wangsiping97/FastGEMV/blob/1fdff6f74aade033c02727a419afd6a4b4bfbc3f/fast_gemv.cu#L14 2. Add `fp16_oss_fast_gemv` to quantize ops benchmark script 3. Add two simple tests for custom op`torch.ops.fbgemm.f16_fast_gemv` to test - `torch.compile()` able - correctness **Next step:** Need fp8 mixed precision support for fast gemv kernel which is what we want Differential Revision: D68470488
YUNQIUGUO
force-pushed
the
export-D68470488
branch
from
b031f69 to
e6c0730
Compare
|
This pull request was exported from Phabricator. Differential Revision: D68470488 |
YUNQIUGUO
force-pushed
the
export-D68470488
branch
from
e6c0730 to
0c02e5a
Compare
YUNQIUGUO
added a commit
to YUNQIUGUO/FBGEMM
that referenced
this pull request
Jan 27, 2025
Summary: X-link: facebookresearch/FBGEMM#688 This diff content includes: 1. Port OSS FastGEMV `fp16` kernel into fbcode and expose to python as a step 1 - `torch.ops.fbgemm.f16_fast_gemv` https://github.com/wangsiping97/FastGEMV/blob/1fdff6f74aade033c02727a419afd6a4b4bfbc3f/fast_gemv.cu#L14 2. Add `fp16_oss_fast_gemv` to quantize ops benchmark script 3. Add two simple tests for custom op`torch.ops.fbgemm.f16_fast_gemv` to test - `torch.compile()` able - correctness **Next step:** Need fp8 mixed precision support for fast gemv kernel which is what we want Differential Revision: D68470488
|
This pull request was exported from Phabricator. Differential Revision: D68470488 |
YUNQIUGUO
force-pushed
the
export-D68470488
branch
from
0c02e5a to
9c2cd76
Compare
YUNQIUGUO
added a commit
to YUNQIUGUO/FBGEMM
that referenced
this pull request
Jan 28, 2025
Summary: X-link: facebookresearch/FBGEMM#688 This diff content includes: 1. Port OSS FastGEMV `fp16` kernel into fbcode and expose to python as a step 1 - `torch.ops.fbgemm.f16_fast_gemv` https://github.com/wangsiping97/FastGEMV/blob/1fdff6f74aade033c02727a419afd6a4b4bfbc3f/fast_gemv.cu#L14 2. Add `fp16_oss_fast_gemv` to quantize ops benchmark script 3. Add two simple tests for custom op`torch.ops.fbgemm.f16_fast_gemv` to test - `torch.compile()` able - correctness Perf numbers: P1720649201 compared with `f16_baseline,fp16_oss_fast_gemv,cuda_lite,marlin_bf16i4,machete_bf16i4` **Next step:** Need fp8 mixed precision support for fast gemv kernel which is what we want Differential Revision: D68470488
|
This pull request was exported from Phabricator. Differential Revision: D68470488 |
YUNQIUGUO
added a commit
to YUNQIUGUO/FBGEMM
that referenced
this pull request
Jan 28, 2025
Summary: X-link: facebookresearch/FBGEMM#688 This diff content includes: 1. Port OSS FastGEMV `fp16` kernel into fbcode and expose to python as a step 1 - `torch.ops.fbgemm.f16_fast_gemv` https://github.com/wangsiping97/FastGEMV/blob/1fdff6f74aade033c02727a419afd6a4b4bfbc3f/fast_gemv.cu#L14 2. Add `fp16_oss_fast_gemv` to quantize ops benchmark script 3. Add two simple tests for custom op`torch.ops.fbgemm.f16_fast_gemv` to test - `torch.compile()` able - correctness Perf numbers: P1720649201 compared with `f16_baseline,fp16_oss_fast_gemv,cuda_lite,marlin_bf16i4,machete_bf16i4` **Next step:** Need fp8 mixed precision support for fast gemv kernel which is what we want Differential Revision: D68470488
YUNQIUGUO
force-pushed
the
export-D68470488
branch
from
9c2cd76 to
6c11e8e
Compare
|
This pull request was exported from Phabricator. Differential Revision: D68470488 |
YUNQIUGUO
force-pushed
the
export-D68470488
branch
from
6c11e8e to
2864803
Compare
YUNQIUGUO
added a commit
to YUNQIUGUO/FBGEMM
that referenced
this pull request
Jan 30, 2025
Summary: X-link: facebookresearch/FBGEMM#688 This diff content includes: 1. Port OSS FastGEMV `fp16` kernel into fbcode and expose to python as a step 1 - `torch.ops.fbgemm.f16_fast_gemv` https://github.com/wangsiping97/FastGEMV/blob/1fdff6f74aade033c02727a419afd6a4b4bfbc3f/fast_gemv.cu#L14 2. Add `fp16_oss_fast_gemv` to quantize ops benchmark script 3. Add two simple tests for custom op`torch.ops.fbgemm.f16_fast_gemv` to test - `torch.compile()` able - correctness Perf numbers: P1722119058 compared with `f16_baseline,fp16_oss_fast_gemv,cuda_lite,marlin_bf16i4,machete_bf16i4` heuristic sweep results from the 4 problem sizes we care about: P1722043272 **Next step:** Need fp8 mixed precision support for fast gemv kernel which is what we want Differential Revision: D68470488
|
This pull request was exported from Phabricator. Differential Revision: D68470488 |
YUNQIUGUO
added a commit
to YUNQIUGUO/FBGEMM
that referenced
this pull request
Jan 30, 2025
Summary: X-link: facebookresearch/FBGEMM#688 This diff content includes: 1. Port OSS FastGEMV `fp16` kernel into fbcode and expose to python as a step 1 - `torch.ops.fbgemm.f16_fast_gemv` https://github.com/wangsiping97/FastGEMV/blob/1fdff6f74aade033c02727a419afd6a4b4bfbc3f/fast_gemv.cu#L14 2. Add `fp16_oss_fast_gemv` to quantize ops benchmark script 3. Add two simple tests for custom op`torch.ops.fbgemm.f16_fast_gemv` to test - `torch.compile()` able - correctness Perf numbers: P1722119058 compared with `f16_baseline,fp16_oss_fast_gemv,cuda_lite,marlin_bf16i4,machete_bf16i4` ====================== ### Benchmark Results | **M** | **N** | **K** | **Method** | **Elapsed Time (ms)** | **TFLOPS** | **GB/s** | | --- | --- | --- | --- | --- | --- | --- | | 1 | 1280 | 8192 | bf16_baseline | 0.024 | 0.860 | 861.042 | | 1 | 1280 | 8192 | fp16_oss_fast_gemv | 0.019 | 1.126 | 1127.391 | | 1 | 1280 | 8192 | cuda_lite | 0.015 | 1.357 | 679.032 | | 1 | 1280 | 8192 | marlin_bf16i4 | 0.027 | 0.768 | 192.612 | | 1 | 1280 | 8192 | machete_bf16i4 | 0.026 | 0.810 | 203.219 | | 1 | 8192 | 1024 | bf16_baseline | 0.018 | 0.952 | 953.176 | | 1 | 8192 | 1024 | fp16_oss_fast_gemv | 0.010 | 1.763 | 1765.033 | | 1 | 8192 | 1024 | cuda_lite | 0.014 | 1.198 | 600.054 | | 1 | 8192 | 1024 | marlin_bf16i4 | 0.015 | 1.144 | 287.150 | | 1 | 8192 | 1024 | machete_bf16i4 | 0.014 | 1.187 | 298.096 | | 1 | 7168 | 8192 | bf16_baseline | 0.073 | 1.609 | 1608.983 | | 1 | 7168 | 8192 | fp16_oss_fast_gemv | 0.069 | 1.697 | 1697.308 | | 1 | 7168 | 8192 | cuda_lite | 0.044 | 2.679 | 1340.093 | | 1 | 7168 | 8192 | marlin_bf16i4 | 0.033 | 3.590 | 898.436 | | 1 | 7168 | 8192 | machete_bf16i4 | 0.039 | 3.017 | 755.147 | | 1 | 8192 | 3584 | bf16_baseline | 0.045 | 1.312 | 1312.239 | | 1 | 8192 | 3584 | fp16_oss_fast_gemv | 0.026 | 2.268 | 1134.843 | | 1 | 8192 | 3584 | cuda_lite | 0.026 | 2.271 | 1136.151 | | 1 | 8192 | 3584 | marlin_bf16i4 | 0.021 | 2.808 | 703.164 | | 1 | 8192 | 3584 | machete_bf16i4 | 0.024 | 2.460 | 615.990 | heuristic sweep results from the 4 problem sizes we care about: P1722043272 **Next step:** Need fp8 mixed precision support for fast gemv kernel which is what we want Differential Revision: D68470488
YUNQIUGUO
force-pushed
the
export-D68470488
branch
from
2864803 to
c79bbc3
Compare
|
This pull request was exported from Phabricator. Differential Revision: D68470488 |
YUNQIUGUO
added a commit
to YUNQIUGUO/FBGEMM
that referenced
this pull request
Jan 31, 2025
Summary: X-link: facebookresearch/FBGEMM#688 This diff content includes: 1. Port OSS FastGEMV `fp16` kernel into fbcode and expose to python as a step 1 - `torch.ops.fbgemm.f16_fast_gemv` https://github.com/wangsiping97/FastGEMV/blob/1fdff6f74aade033c02727a419afd6a4b4bfbc3f/fast_gemv.cu#L14 2. Add `fp16_oss_fast_gemv` to quantize ops benchmark script 3. Add two simple tests for custom op`torch.ops.fbgemm.f16_fast_gemv` to test - `torch.compile()` able - correctness Perf numbers compared with `f16_baseline,fp16_oss_fast_gemv,cuda_lite,marlin_bf16i4,machete_bf16i4` ====================== ### Benchmark Results | **M** | **N** | **K** | **Method** | **Elapsed Time (ms)** | **TFLOPS** | **GB/s** | | --- | --- | --- | --- | --- | --- | --- | | 1 | 1280 | 8192 | bf16_baseline | 0.024 | 0.860 | 861.042 | | 1 | 1280 | 8192 | fp16_oss_fast_gemv | 0.019 | 1.126 | 1127.391 | | 1 | 1280 | 8192 | cuda_lite_fp8 | 0.015 | 1.357 | 679.032 | | 1 | 1280 | 8192 | marlin_bf16i4 | 0.027 | 0.768 | 192.612 | | 1 | 1280 | 8192 | machete_bf16i4 | 0.026 | 0.810 | 203.219 | | 1 | 8192 | 1024 | bf16_baseline | 0.018 | 0.952 | 953.176 | | 1 | 8192 | 1024 | fp16_oss_fast_gemv | 0.010 | 1.763 | 1765.033 | | 1 | 8192 | 1024 | cuda_lite_fp8 | 0.014 | 1.198 | 600.054 | | 1 | 8192 | 1024 | marlin_bf16i4 | 0.015 | 1.144 | 287.150 | | 1 | 8192 | 1024 | machete_bf16i4 | 0.014 | 1.187 | 298.096 | | 1 | 7168 | 8192 | bf16_baseline | 0.073 | 1.609 | 1608.983 | | 1 | 7168 | 8192 | fp16_oss_fast_gemv | 0.069 | 1.697 | 1697.308 | | 1 | 7168 | 8192 | cuda_lite_fp8 | 0.044 | 2.679 | 1340.093 | | 1 | 7168 | 8192 | marlin_bf16i4 | 0.033 | 3.590 | 898.436 | | 1 | 7168 | 8192 | machete_bf16i4 | 0.039 | 3.017 | 755.147 | | 1 | 8192 | 3584 | bf16_baseline | 0.045 | 1.312 | 1312.239 | | 1 | 8192 | 3584 | fp16_oss_fast_gemv | 0.026 | 2.268 | 1134.843 | | 1 | 8192 | 3584 | cuda_lite_fp8 | 0.026 | 2.271 | 1136.151 | | 1 | 8192 | 3584 | marlin_bf16i4 | 0.021 | 2.808 | 703.164 | | 1 | 8192 | 3584 | machete_bf16i4 | 0.024 | 2.460 | 615.990 | Note that currently the precision with fast_gemv kernel and cuda_lite not matched yet. so need fp8 to coming in for a fairer result. heuristic sweep results from the 4 problem sizes we care about: P1722806148 **Next step:** Need fp8 mixed precision support for fast gemv kernel which is what we want Differential Revision: D68470488
YUNQIUGUO
force-pushed
the
export-D68470488
branch
from
c79bbc3 to
86df5d1
Compare
|
This pull request was exported from Phabricator. Differential Revision: D68470488 |
YUNQIUGUO
added a commit
to YUNQIUGUO/FBGEMM
that referenced
this pull request
Jan 31, 2025
Summary: X-link: facebookresearch/FBGEMM#688 This diff content includes: 1. Port OSS FastGEMV `fp16` kernel into fbcode and expose to python as a step 1 - `torch.ops.fbgemm.f16_fast_gemv` https://github.com/wangsiping97/FastGEMV/blob/1fdff6f74aade033c02727a419afd6a4b4bfbc3f/fast_gemv.cu#L14 2. Add `fp16_oss_fast_gemv` to quantize ops benchmark script 3. Add two simple tests for custom op`torch.ops.fbgemm.f16_fast_gemv` to test - `torch.compile()` able - correctness Perf numbers compared with `f16_baseline,fp16_oss_fast_gemv,cuda_lite,marlin_bf16i4,machete_bf16i4` ====================== ### Benchmark Results | **M** | **N** | **K** | **Method** | **Elapsed Time (ms)** | **TFLOPS** | **GB/s** | | --- | --- | --- | --- | --- | --- | --- | | 1 | 1280 | 8192 | bf16_baseline | 0.024 | 0.860 | 861.042 | | 1 | 1280 | 8192 | fp16_oss_fast_gemv | 0.019 | 1.126 | 1127.391 | | 1 | 1280 | 8192 | cuda_lite_fp8 | 0.015 | 1.357 | 679.032 | | 1 | 1280 | 8192 | marlin_bf16i4 | 0.027 | 0.768 | 192.612 | | 1 | 1280 | 8192 | machete_bf16i4 | 0.026 | 0.810 | 203.219 | | 1 | 8192 | 1024 | bf16_baseline | 0.018 | 0.952 | 953.176 | | 1 | 8192 | 1024 | fp16_oss_fast_gemv | 0.010 | 1.763 | 1765.033 | | 1 | 8192 | 1024 | cuda_lite_fp8 | 0.014 | 1.198 | 600.054 | | 1 | 8192 | 1024 | marlin_bf16i4 | 0.015 | 1.144 | 287.150 | | 1 | 8192 | 1024 | machete_bf16i4 | 0.014 | 1.187 | 298.096 | | 1 | 7168 | 8192 | bf16_baseline | 0.073 | 1.609 | 1608.983 | | 1 | 7168 | 8192 | fp16_oss_fast_gemv | 0.069 | 1.697 | 1697.308 | | 1 | 7168 | 8192 | cuda_lite_fp8 | 0.044 | 2.679 | 1340.093 | | 1 | 7168 | 8192 | marlin_bf16i4 | 0.033 | 3.590 | 898.436 | | 1 | 7168 | 8192 | machete_bf16i4 | 0.039 | 3.017 | 755.147 | | 1 | 8192 | 3584 | bf16_baseline | 0.045 | 1.312 | 1312.239 | | 1 | 8192 | 3584 | fp16_oss_fast_gemv | 0.026 | 2.268 | 1134.843 | | 1 | 8192 | 3584 | cuda_lite_fp8 | 0.026 | 2.271 | 1136.151 | | 1 | 8192 | 3584 | marlin_bf16i4 | 0.021 | 2.808 | 703.164 | | 1 | 8192 | 3584 | machete_bf16i4 | 0.024 | 2.460 | 615.990 | Note that currently the precision with fast_gemv kernel and cuda_lite not matched yet. so need fp8 to coming in for a fairer result. heuristic sweep results from the 4 problem sizes we care about: P1722806148 **Next step:** Need fp8 mixed precision support for fast gemv kernel which is what we want Differential Revision: D68470488
YUNQIUGUO
force-pushed
the
export-D68470488
branch
from
86df5d1 to
949bdfc
Compare
YUNQIUGUO
added a commit
to YUNQIUGUO/FBGEMM
that referenced
this pull request
Feb 4, 2025
Summary: X-link: facebookresearch/FBGEMM#688 This diff content includes: 1. Port OSS FastGEMV `bf16` kernel into fbcode and expose to python as a step 1 - `torch.ops.fbgemm.bf16_fast_gemv` https://github.com/wangsiping97/FastGEMV/blob/1fdff6f74aade033c02727a419afd6a4b4bfbc3f/fast_gemv.cu#L14 2. Add `bf16_oss_fast_gemv` to quantize ops benchmark script 3. Add two simple tests for custom op`torch.ops.fbgemm.f16_fast_gemv` to test - `torch.compile()` able - correctness Perf numbers compared with `bf16_baseline,bf16_oss_fast_gemv,cuda_lite,marlin_bf16i4,machete_bf16i4` ====================== ### Benchmark Results on H100 | **M** | **N** | **K** | **Method** | **Elapsed Time (ms)** | **TFLOPS** | **GB/s** | | --- | --- | --- | --- | --- | --- | --- | | 1 | 1280 | 8192 | bf16_baseline | 0.024 | 0.860 | 861.042 | | 1 | 1280 | 8192 | bf16_oss_fast_gemv | 0.019 | 1.126 | 1127.391 | | 1 | 1280 | 8192 | cuda_lite_fp8 | 0.015 | 1.357 | 679.032 | | 1 | 1280 | 8192 | marlin_bf16i4 | 0.027 | 0.768 | 192.612 | | 1 | 1280 | 8192 | machete_bf16i4 | 0.026 | 0.810 | 203.219 | | 1 | 8192 | 1024 | bf16_baseline | 0.018 | 0.952 | 953.176 | | 1 | 8192 | 1024 | bf16_oss_fast_gemv | 0.015 | 1.100 | 1100.900 | | 1 | 8192 | 1024 | cuda_lite_fp8 | 0.014 | 1.198 | 600.054 | | 1 | 8192 | 1024 | marlin_bf16i4 | 0.015 | 1.144 | 287.150 | | 1 | 8192 | 1024 | machete_bf16i4 | 0.014 | 1.187 | 298.096 | | 1 | 7168 | 8192 | bf16_baseline | 0.073 | 1.609 | 1608.983 | | 1 | 7168 | 8192 | bf16_oss_fast_gemv | 0.069 | 1.697 | 1697.308 | | 1 | 7168 | 8192 | cuda_lite_fp8 | 0.044 | 2.679 | 1340.093 | | 1 | 7168 | 8192 | marlin_bf16i4 | 0.033 | 3.590 | 898.436 | | 1 | 7168 | 8192 | machete_bf16i4 | 0.039 | 3.017 | 755.147 | | 1 | 8192 | 3584 | bf16_baseline | 0.045 | 1.312 | 1312.239 | | 1 | 8192 | 3584 | bf16_oss_fast_gemv | 0.041 | 1.427 | 1427.166 | | 1 | 8192 | 3584 | cuda_lite_fp8 | 0.026 | 2.271 | 1136.151 | | 1 | 8192 | 3584 | marlin_bf16i4 | 0.021 | 2.808 | 703.164 | | 1 | 8192 | 3584 | machete_bf16i4 | 0.024 | 2.460 | 615.990 | Note that currently the precision with `fast_gemv` kernel and `cuda_lite` does not match yet. so need fp8 to coming in for a fairer result. Also no_cuda_graph flag enabled when running the quantize_bench heuristic sweep results from the 4 problem sizes we care about: P1722806148 **Next step:** Need fp8 mixed precision support for fast gemv kernel which is what we want Differential Revision: D68470488
|
This pull request was exported from Phabricator. Differential Revision: D68470488 |
YUNQIUGUO
added a commit
to YUNQIUGUO/FBGEMM
that referenced
this pull request
Feb 4, 2025
Summary: X-link: facebookresearch/FBGEMM#688 This diff content includes: 1. Port OSS FastGEMV `bf16` kernel into fbcode and expose to python as a step 1 - `torch.ops.fbgemm.bf16_fast_gemv` https://github.com/wangsiping97/FastGEMV/blob/1fdff6f74aade033c02727a419afd6a4b4bfbc3f/fast_gemv.cu#L14 2. Add `bf16_oss_fast_gemv` to quantize ops benchmark script 3. Add two simple tests for custom op`torch.ops.fbgemm.f16_fast_gemv` to test - `torch.compile()` able - correctness Perf numbers compared with `bf16_baseline,bf16_oss_fast_gemv,cuda_lite,marlin_bf16i4,machete_bf16i4` ====================== ### Benchmark Results on H100 | **M** | **N** | **K** | **Method** | **Elapsed Time (ms)** | **TFLOPS** | **GB/s** | | --- | --- | --- | --- | --- | --- | --- | | 1 | 1280 | 8192 | bf16_baseline | 0.024 | 0.860 | 861.042 | | 1 | 1280 | 8192 | bf16_oss_fast_gemv | 0.019 | 1.126 | 1127.391 | | 1 | 1280 | 8192 | cuda_lite_fp8 | 0.015 | 1.357 | 679.032 | | 1 | 1280 | 8192 | marlin_bf16i4 | 0.027 | 0.768 | 192.612 | | 1 | 1280 | 8192 | machete_bf16i4 | 0.026 | 0.810 | 203.219 | | 1 | 8192 | 1024 | bf16_baseline | 0.018 | 0.952 | 953.176 | | 1 | 8192 | 1024 | bf16_oss_fast_gemv | 0.015 | 1.100 | 1100.900 | | 1 | 8192 | 1024 | cuda_lite_fp8 | 0.014 | 1.198 | 600.054 | | 1 | 8192 | 1024 | marlin_bf16i4 | 0.015 | 1.144 | 287.150 | | 1 | 8192 | 1024 | machete_bf16i4 | 0.014 | 1.187 | 298.096 | | 1 | 7168 | 8192 | bf16_baseline | 0.073 | 1.609 | 1608.983 | | 1 | 7168 | 8192 | bf16_oss_fast_gemv | 0.069 | 1.697 | 1697.308 | | 1 | 7168 | 8192 | cuda_lite_fp8 | 0.044 | 2.679 | 1340.093 | | 1 | 7168 | 8192 | marlin_bf16i4 | 0.033 | 3.590 | 898.436 | | 1 | 7168 | 8192 | machete_bf16i4 | 0.039 | 3.017 | 755.147 | | 1 | 8192 | 3584 | bf16_baseline | 0.045 | 1.312 | 1312.239 | | 1 | 8192 | 3584 | bf16_oss_fast_gemv | 0.041 | 1.427 | 1427.166 | | 1 | 8192 | 3584 | cuda_lite_fp8 | 0.026 | 2.271 | 1136.151 | | 1 | 8192 | 3584 | marlin_bf16i4 | 0.021 | 2.808 | 703.164 | | 1 | 8192 | 3584 | machete_bf16i4 | 0.024 | 2.460 | 615.990 | Note that currently the precision with `fast_gemv` kernel and `cuda_lite` does not match yet. so need fp8 to coming in for a fairer result. Also no_cuda_graph flag enabled when running the quantize_bench heuristic sweep results from the 4 problem sizes we care about: P1722806148 **Next step:** Need fp8 mixed precision support for fast gemv kernel which is what we want Differential Revision: D68470488
YUNQIUGUO
force-pushed
the
export-D68470488
branch
from
876d9c3 to
5de93cd
Compare
|
This pull request was exported from Phabricator. Differential Revision: D68470488 |
YUNQIUGUO
force-pushed
the
export-D68470488
branch
from
5de93cd to
7104abc
Compare
YUNQIUGUO
added a commit
to YUNQIUGUO/FBGEMM
that referenced
this pull request
Feb 4, 2025
Summary: X-link: facebookresearch/FBGEMM#688 This diff content includes: 1. Port OSS FastGEMV `bf16` kernel into fbcode and expose to python as a step 1 - `torch.ops.fbgemm.bf16_fast_gemv` https://github.com/wangsiping97/FastGEMV/blob/1fdff6f74aade033c02727a419afd6a4b4bfbc3f/fast_gemv.cu#L14 2. Add `bf16_oss_fast_gemv` to quantize ops benchmark script 3. Add two simple tests for custom op`torch.ops.fbgemm.f16_fast_gemv` to test - `torch.compile()` able - correctness Perf numbers compared with `bf16_baseline,bf16_oss_fast_gemv,cuda_lite,marlin_bf16i4,machete_bf16i4` ====================== ### Benchmark Results on H100 | **M** | **N** | **K** | **Method** | **Elapsed Time (ms)** | **TFLOPS** | **GB/s** | | --- | --- | --- | --- | --- | --- | --- | | 1 | 1280 | 8192 | bf16_baseline | 0.024 | 0.860 | 861.042 | | 1 | 1280 | 8192 | bf16_oss_fast_gemv | 0.019 | 1.126 | 1127.391 | | 1 | 1280 | 8192 | cuda_lite_fp8 | 0.015 | 1.357 | 679.032 | | 1 | 1280 | 8192 | marlin_bf16i4 | 0.027 | 0.768 | 192.612 | | 1 | 1280 | 8192 | machete_bf16i4 | 0.026 | 0.810 | 203.219 | | 1 | 8192 | 1024 | bf16_baseline | 0.018 | 0.952 | 953.176 | | 1 | 8192 | 1024 | bf16_oss_fast_gemv | 0.015 | 1.100 | 1100.900 | | 1 | 8192 | 1024 | cuda_lite_fp8 | 0.014 | 1.198 | 600.054 | | 1 | 8192 | 1024 | marlin_bf16i4 | 0.015 | 1.144 | 287.150 | | 1 | 8192 | 1024 | machete_bf16i4 | 0.014 | 1.187 | 298.096 | | 1 | 7168 | 8192 | bf16_baseline | 0.073 | 1.609 | 1608.983 | | 1 | 7168 | 8192 | bf16_oss_fast_gemv | 0.069 | 1.697 | 1697.308 | | 1 | 7168 | 8192 | cuda_lite_fp8 | 0.044 | 2.679 | 1340.093 | | 1 | 7168 | 8192 | marlin_bf16i4 | 0.033 | 3.590 | 898.436 | | 1 | 7168 | 8192 | machete_bf16i4 | 0.039 | 3.017 | 755.147 | | 1 | 8192 | 3584 | bf16_baseline | 0.045 | 1.312 | 1312.239 | | 1 | 8192 | 3584 | bf16_oss_fast_gemv | 0.041 | 1.427 | 1427.166 | | 1 | 8192 | 3584 | cuda_lite_fp8 | 0.026 | 2.271 | 1136.151 | | 1 | 8192 | 3584 | marlin_bf16i4 | 0.021 | 2.808 | 703.164 | | 1 | 8192 | 3584 | machete_bf16i4 | 0.024 | 2.460 | 615.990 | Note that currently the precision with `fast_gemv` kernel and `cuda_lite` does not match yet. so need fp8 to coming in for a fairer result. Also no_cuda_graph flag enabled when running the quantize_bench heuristic sweep results from the 4 problem sizes we care about: P1722806148 **Next step:** Need fp8 mixed precision support for fast gemv kernel which is what we want Differential Revision: D68470488
|
This pull request was exported from Phabricator. Differential Revision: D68470488 |
YUNQIUGUO
added a commit
to YUNQIUGUO/FBGEMM
that referenced
this pull request
Feb 4, 2025
Summary: X-link: facebookresearch/FBGEMM#688 This diff content includes: 1. Port OSS FastGEMV `bf16` kernel into fbcode and expose to python as a step 1 - `torch.ops.fbgemm.bf16_fast_gemv` https://github.com/wangsiping97/FastGEMV/blob/1fdff6f74aade033c02727a419afd6a4b4bfbc3f/fast_gemv.cu#L14 2. Add `bf16_oss_fast_gemv` to quantize ops benchmark script 3. Add two simple tests for custom op`torch.ops.fbgemm.f16_fast_gemv` to test - `torch.compile()` able - correctness Perf numbers compared with `bf16_baseline,bf16_oss_fast_gemv,cuda_lite,marlin_bf16i4,machete_bf16i4` ====================== ### Benchmark Results on H100 | **M** | **N** | **K** | **Method** | **Elapsed Time (ms)** | **TFLOPS** | **GB/s** | | --- | --- | --- | --- | --- | --- | --- | | 1 | 1280 | 8192 | bf16_baseline | 0.024 | 0.860 | 861.042 | | 1 | 1280 | 8192 | bf16_oss_fast_gemv | 0.019 | 1.126 | 1127.391 | | 1 | 1280 | 8192 | cuda_lite_fp8 | 0.015 | 1.357 | 679.032 | | 1 | 1280 | 8192 | marlin_bf16i4 | 0.027 | 0.768 | 192.612 | | 1 | 1280 | 8192 | machete_bf16i4 | 0.026 | 0.810 | 203.219 | | 1 | 8192 | 1024 | bf16_baseline | 0.018 | 0.952 | 953.176 | | 1 | 8192 | 1024 | bf16_oss_fast_gemv | 0.015 | 1.100 | 1100.900 | | 1 | 8192 | 1024 | cuda_lite_fp8 | 0.014 | 1.198 | 600.054 | | 1 | 8192 | 1024 | marlin_bf16i4 | 0.015 | 1.144 | 287.150 | | 1 | 8192 | 1024 | machete_bf16i4 | 0.014 | 1.187 | 298.096 | | 1 | 7168 | 8192 | bf16_baseline | 0.073 | 1.609 | 1608.983 | | 1 | 7168 | 8192 | bf16_oss_fast_gemv | 0.069 | 1.697 | 1697.308 | | 1 | 7168 | 8192 | cuda_lite_fp8 | 0.044 | 2.679 | 1340.093 | | 1 | 7168 | 8192 | marlin_bf16i4 | 0.033 | 3.590 | 898.436 | | 1 | 7168 | 8192 | machete_bf16i4 | 0.039 | 3.017 | 755.147 | | 1 | 8192 | 3584 | bf16_baseline | 0.045 | 1.312 | 1312.239 | | 1 | 8192 | 3584 | bf16_oss_fast_gemv | 0.041 | 1.427 | 1427.166 | | 1 | 8192 | 3584 | cuda_lite_fp8 | 0.026 | 2.271 | 1136.151 | | 1 | 8192 | 3584 | marlin_bf16i4 | 0.021 | 2.808 | 703.164 | | 1 | 8192 | 3584 | machete_bf16i4 | 0.024 | 2.460 | 615.990 | Note that currently the precision with `fast_gemv` kernel and `cuda_lite` does not match yet. so need fp8 to coming in for a fairer result. Also no_cuda_graph flag enabled when running the quantize_bench heuristic sweep results from the 4 problem sizes we care about: P1722806148 **Next step:** Need fp8 mixed precision support for fast gemv kernel which is what we want Differential Revision: D68470488
YUNQIUGUO
force-pushed
the
export-D68470488
branch
from
7104abc to
57b68a3
Compare
|
This pull request was exported from Phabricator. Differential Revision: D68470488 |
YUNQIUGUO
added a commit
to YUNQIUGUO/FBGEMM
that referenced
this pull request
Feb 4, 2025
Summary: X-link: facebookresearch/FBGEMM#688 This diff content includes: 1. Port OSS FastGEMV `bf16` kernel into fbcode and expose to python as a step 1 - `torch.ops.fbgemm.bf16_fast_gemv` https://github.com/wangsiping97/FastGEMV/blob/1fdff6f74aade033c02727a419afd6a4b4bfbc3f/fast_gemv.cu#L14 2. Add `bf16_oss_fast_gemv` to quantize ops benchmark script 3. Add two simple tests for custom op`torch.ops.fbgemm.f16_fast_gemv` to test - `torch.compile()` able - correctness Perf numbers compared with `bf16_baseline,bf16_oss_fast_gemv,cuda_lite,marlin_bf16i4,machete_bf16i4` ====================== ### Benchmark Results on H100 | **M** | **N** | **K** | **Method** | **Elapsed Time (ms)** | **TFLOPS** | **GB/s** | | --- | --- | --- | --- | --- | --- | --- | | 1 | 1280 | 8192 | bf16_baseline | 0.024 | 0.860 | 861.042 | | 1 | 1280 | 8192 | bf16_oss_fast_gemv | 0.019 | 1.126 | 1127.391 | | 1 | 1280 | 8192 | cuda_lite_fp8 | 0.015 | 1.357 | 679.032 | | 1 | 1280 | 8192 | marlin_bf16i4 | 0.027 | 0.768 | 192.612 | | 1 | 1280 | 8192 | machete_bf16i4 | 0.026 | 0.810 | 203.219 | | 1 | 8192 | 1024 | bf16_baseline | 0.018 | 0.952 | 953.176 | | 1 | 8192 | 1024 | bf16_oss_fast_gemv | 0.015 | 1.100 | 1100.900 | | 1 | 8192 | 1024 | cuda_lite_fp8 | 0.014 | 1.198 | 600.054 | | 1 | 8192 | 1024 | marlin_bf16i4 | 0.015 | 1.144 | 287.150 | | 1 | 8192 | 1024 | machete_bf16i4 | 0.014 | 1.187 | 298.096 | | 1 | 7168 | 8192 | bf16_baseline | 0.073 | 1.609 | 1608.983 | | 1 | 7168 | 8192 | bf16_oss_fast_gemv | 0.069 | 1.697 | 1697.308 | | 1 | 7168 | 8192 | cuda_lite_fp8 | 0.044 | 2.679 | 1340.093 | | 1 | 7168 | 8192 | marlin_bf16i4 | 0.033 | 3.590 | 898.436 | | 1 | 7168 | 8192 | machete_bf16i4 | 0.039 | 3.017 | 755.147 | | 1 | 8192 | 3584 | bf16_baseline | 0.045 | 1.312 | 1312.239 | | 1 | 8192 | 3584 | bf16_oss_fast_gemv | 0.041 | 1.427 | 1427.166 | | 1 | 8192 | 3584 | cuda_lite_fp8 | 0.026 | 2.271 | 1136.151 | | 1 | 8192 | 3584 | marlin_bf16i4 | 0.021 | 2.808 | 703.164 | | 1 | 8192 | 3584 | machete_bf16i4 | 0.024 | 2.460 | 615.990 | Note that currently the precision with `fast_gemv` kernel and `cuda_lite` does not match yet. so need fp8 to coming in for a fairer result. Also no_cuda_graph flag enabled when running the quantize_bench heuristic sweep results from the 4 problem sizes we care about: P1722806148 **Next step:** Need fp8 mixed precision support for fast gemv kernel which is what we want Differential Revision: D68470488
YUNQIUGUO
force-pushed
the
export-D68470488
branch
from
57b68a3 to
f8bcdd6
Compare
|
This pull request was exported from Phabricator. Differential Revision: D68470488 |
YUNQIUGUO
added a commit
to YUNQIUGUO/FBGEMM
that referenced
this pull request
Feb 4, 2025
Summary: X-link: facebookresearch/FBGEMM#688 This diff content includes: 1. Port OSS FastGEMV `bf16` kernel into fbcode and expose to python as a step 1 - `torch.ops.fbgemm.bf16_fast_gemv` https://github.com/wangsiping97/FastGEMV/blob/1fdff6f74aade033c02727a419afd6a4b4bfbc3f/fast_gemv.cu#L14 2. Add `bf16_oss_fast_gemv` to quantize ops benchmark script 3. Add two simple tests for custom op`torch.ops.fbgemm.f16_fast_gemv` to test - `torch.compile()` able - correctness Perf numbers compared with `bf16_baseline,bf16_oss_fast_gemv,cuda_lite,marlin_bf16i4,machete_bf16i4` ====================== ### Benchmark Results on H100 | **M** | **N** | **K** | **Method** | **Elapsed Time (ms)** | **TFLOPS** | **GB/s** | | --- | --- | --- | --- | --- | --- | --- | | 1 | 1280 | 8192 | bf16_baseline | 0.024 | 0.860 | 861.042 | | 1 | 1280 | 8192 | bf16_oss_fast_gemv | 0.019 | 1.126 | 1127.391 | | 1 | 1280 | 8192 | cuda_lite_fp8 | 0.015 | 1.357 | 679.032 | | 1 | 1280 | 8192 | marlin_bf16i4 | 0.027 | 0.768 | 192.612 | | 1 | 1280 | 8192 | machete_bf16i4 | 0.026 | 0.810 | 203.219 | | 1 | 8192 | 1024 | bf16_baseline | 0.018 | 0.952 | 953.176 | | 1 | 8192 | 1024 | bf16_oss_fast_gemv | 0.015 | 1.100 | 1100.900 | | 1 | 8192 | 1024 | cuda_lite_fp8 | 0.014 | 1.198 | 600.054 | | 1 | 8192 | 1024 | marlin_bf16i4 | 0.015 | 1.144 | 287.150 | | 1 | 8192 | 1024 | machete_bf16i4 | 0.014 | 1.187 | 298.096 | | 1 | 7168 | 8192 | bf16_baseline | 0.073 | 1.609 | 1608.983 | | 1 | 7168 | 8192 | bf16_oss_fast_gemv | 0.069 | 1.697 | 1697.308 | | 1 | 7168 | 8192 | cuda_lite_fp8 | 0.044 | 2.679 | 1340.093 | | 1 | 7168 | 8192 | marlin_bf16i4 | 0.033 | 3.590 | 898.436 | | 1 | 7168 | 8192 | machete_bf16i4 | 0.039 | 3.017 | 755.147 | | 1 | 8192 | 3584 | bf16_baseline | 0.045 | 1.312 | 1312.239 | | 1 | 8192 | 3584 | bf16_oss_fast_gemv | 0.041 | 1.427 | 1427.166 | | 1 | 8192 | 3584 | cuda_lite_fp8 | 0.026 | 2.271 | 1136.151 | | 1 | 8192 | 3584 | marlin_bf16i4 | 0.021 | 2.808 | 703.164 | | 1 | 8192 | 3584 | machete_bf16i4 | 0.024 | 2.460 | 615.990 | Note that currently the precision with `fast_gemv` kernel and `cuda_lite` does not match yet. so need fp8 to coming in for a fairer result. Also no_cuda_graph flag enabled when running the quantize_bench heuristic sweep results from the 4 problem sizes we care about: P1722806148 **Next step:** Need fp8 mixed precision support for fast gemv kernel which is what we want Differential Revision: D68470488
YUNQIUGUO
force-pushed
the
export-D68470488
branch
from
f8bcdd6 to
f348dc9
Compare
|
This pull request was exported from Phabricator. Differential Revision: D68470488 |
YUNQIUGUO
added a commit
to YUNQIUGUO/FBGEMM
that referenced
this pull request
Feb 4, 2025
Summary: X-link: facebookresearch/FBGEMM#688 This diff content includes: 1. Port OSS FastGEMV `bf16` kernel into fbcode and expose to python as a step 1 - `torch.ops.fbgemm.bf16_fast_gemv` https://github.com/wangsiping97/FastGEMV/blob/1fdff6f74aade033c02727a419afd6a4b4bfbc3f/fast_gemv.cu#L14 2. Add `bf16_oss_fast_gemv` to quantize ops benchmark script 3. Add two simple tests for custom op`torch.ops.fbgemm.f16_fast_gemv` to test - `torch.compile()` able - correctness Perf numbers compared with `bf16_baseline,bf16_oss_fast_gemv,cuda_lite,marlin_bf16i4,machete_bf16i4` ====================== ### Benchmark Results on H100 | **M** | **N** | **K** | **Method** | **Elapsed Time (ms)** | **TFLOPS** | **GB/s** | | --- | --- | --- | --- | --- | --- | --- | | 1 | 1280 | 8192 | bf16_baseline | 0.024 | 0.860 | 861.042 | | 1 | 1280 | 8192 | bf16_oss_fast_gemv | 0.019 | 1.126 | 1127.391 | | 1 | 1280 | 8192 | cuda_lite_fp8 | 0.015 | 1.357 | 679.032 | | 1 | 1280 | 8192 | marlin_bf16i4 | 0.027 | 0.768 | 192.612 | | 1 | 1280 | 8192 | machete_bf16i4 | 0.026 | 0.810 | 203.219 | | 1 | 8192 | 1024 | bf16_baseline | 0.018 | 0.952 | 953.176 | | 1 | 8192 | 1024 | bf16_oss_fast_gemv | 0.015 | 1.100 | 1100.900 | | 1 | 8192 | 1024 | cuda_lite_fp8 | 0.014 | 1.198 | 600.054 | | 1 | 8192 | 1024 | marlin_bf16i4 | 0.015 | 1.144 | 287.150 | | 1 | 8192 | 1024 | machete_bf16i4 | 0.014 | 1.187 | 298.096 | | 1 | 7168 | 8192 | bf16_baseline | 0.073 | 1.609 | 1608.983 | | 1 | 7168 | 8192 | bf16_oss_fast_gemv | 0.069 | 1.697 | 1697.308 | | 1 | 7168 | 8192 | cuda_lite_fp8 | 0.044 | 2.679 | 1340.093 | | 1 | 7168 | 8192 | marlin_bf16i4 | 0.033 | 3.590 | 898.436 | | 1 | 7168 | 8192 | machete_bf16i4 | 0.039 | 3.017 | 755.147 | | 1 | 8192 | 3584 | bf16_baseline | 0.045 | 1.312 | 1312.239 | | 1 | 8192 | 3584 | bf16_oss_fast_gemv | 0.041 | 1.427 | 1427.166 | | 1 | 8192 | 3584 | cuda_lite_fp8 | 0.026 | 2.271 | 1136.151 | | 1 | 8192 | 3584 | marlin_bf16i4 | 0.021 | 2.808 | 703.164 | | 1 | 8192 | 3584 | machete_bf16i4 | 0.024 | 2.460 | 615.990 | Note that currently the precision with `fast_gemv` kernel and `cuda_lite` does not match yet. so need fp8 to coming in for a fairer result. Also no_cuda_graph flag enabled when running the quantize_bench heuristic sweep results from the 4 problem sizes we care about: P1722806148 **Next step:** Need fp8 mixed precision support for fast gemv kernel which is what we want Differential Revision: D68470488
YUNQIUGUO
added a commit
to YUNQIUGUO/FBGEMM
that referenced
this pull request
Feb 4, 2025
Summary: X-link: facebookresearch/FBGEMM#688 This diff content includes: 1. Port OSS FastGEMV `bf16` kernel into fbcode and expose to python as a step 1 - `torch.ops.fbgemm.bf16_fast_gemv` https://github.com/wangsiping97/FastGEMV/blob/1fdff6f74aade033c02727a419afd6a4b4bfbc3f/fast_gemv.cu#L14 2. Add `bf16_oss_fast_gemv` to quantize ops benchmark script 3. Add two simple tests for custom op`torch.ops.fbgemm.f16_fast_gemv` to test - `torch.compile()` able - correctness Perf numbers compared with `bf16_baseline,bf16_oss_fast_gemv,cuda_lite,marlin_bf16i4,machete_bf16i4` ====================== ### Benchmark Results on H100 | **M** | **N** | **K** | **Method** | **Elapsed Time (ms)** | **TFLOPS** | **GB/s** | | --- | --- | --- | --- | --- | --- | --- | | 1 | 1280 | 8192 | bf16_baseline | 0.024 | 0.860 | 861.042 | | 1 | 1280 | 8192 | bf16_oss_fast_gemv | 0.019 | 1.126 | 1127.391 | | 1 | 1280 | 8192 | cuda_lite_fp8 | 0.015 | 1.357 | 679.032 | | 1 | 1280 | 8192 | marlin_bf16i4 | 0.027 | 0.768 | 192.612 | | 1 | 1280 | 8192 | machete_bf16i4 | 0.026 | 0.810 | 203.219 | | 1 | 8192 | 1024 | bf16_baseline | 0.018 | 0.952 | 953.176 | | 1 | 8192 | 1024 | bf16_oss_fast_gemv | 0.015 | 1.100 | 1100.900 | | 1 | 8192 | 1024 | cuda_lite_fp8 | 0.014 | 1.198 | 600.054 | | 1 | 8192 | 1024 | marlin_bf16i4 | 0.015 | 1.144 | 287.150 | | 1 | 8192 | 1024 | machete_bf16i4 | 0.014 | 1.187 | 298.096 | | 1 | 7168 | 8192 | bf16_baseline | 0.073 | 1.609 | 1608.983 | | 1 | 7168 | 8192 | bf16_oss_fast_gemv | 0.069 | 1.697 | 1697.308 | | 1 | 7168 | 8192 | cuda_lite_fp8 | 0.044 | 2.679 | 1340.093 | | 1 | 7168 | 8192 | marlin_bf16i4 | 0.033 | 3.590 | 898.436 | | 1 | 7168 | 8192 | machete_bf16i4 | 0.039 | 3.017 | 755.147 | | 1 | 8192 | 3584 | bf16_baseline | 0.045 | 1.312 | 1312.239 | | 1 | 8192 | 3584 | bf16_oss_fast_gemv | 0.041 | 1.427 | 1427.166 | | 1 | 8192 | 3584 | cuda_lite_fp8 | 0.026 | 2.271 | 1136.151 | | 1 | 8192 | 3584 | marlin_bf16i4 | 0.021 | 2.808 | 703.164 | | 1 | 8192 | 3584 | machete_bf16i4 | 0.024 | 2.460 | 615.990 | Note that currently the precision with `fast_gemv` kernel and `cuda_lite` does not match yet. so need fp8 to coming in for a fairer result. Also no_cuda_graph flag enabled when running the quantize_bench heuristic sweep results from the 4 problem sizes we care about: P1722806148 **Next step:** Need fp8 mixed precision support for fast gemv kernel which is what we want Differential Revision: D68470488
YUNQIUGUO
force-pushed
the
export-D68470488
branch
from
f348dc9 to
d3aecb7
Compare
|
This pull request was exported from Phabricator. Differential Revision: D68470488 |
YUNQIUGUO
added a commit
to YUNQIUGUO/FBGEMM
that referenced
this pull request
Feb 4, 2025
Summary: X-link: facebookresearch/FBGEMM#688 This diff content includes: 1. Port OSS FastGEMV `bf16` kernel into fbcode and expose to python as a step 1 - `torch.ops.fbgemm.bf16_fast_gemv` https://github.com/wangsiping97/FastGEMV/blob/1fdff6f74aade033c02727a419afd6a4b4bfbc3f/fast_gemv.cu#L14 2. Add `bf16_oss_fast_gemv` to quantize ops benchmark script 3. Add two simple tests for custom op`torch.ops.fbgemm.f16_fast_gemv` to test - `torch.compile()` able - correctness Perf numbers compared with `bf16_baseline,bf16_oss_fast_gemv,cuda_lite,marlin_bf16i4,machete_bf16i4` ====================== ### Benchmark Results on H100 | **M** | **N** | **K** | **Method** | **Elapsed Time (ms)** | **TFLOPS** | **GB/s** | | --- | --- | --- | --- | --- | --- | --- | | 1 | 1280 | 8192 | bf16_baseline | 0.024 | 0.860 | 861.042 | | 1 | 1280 | 8192 | bf16_oss_fast_gemv | 0.019 | 1.126 | 1127.391 | | 1 | 1280 | 8192 | cuda_lite_fp8 | 0.015 | 1.357 | 679.032 | | 1 | 1280 | 8192 | marlin_bf16i4 | 0.027 | 0.768 | 192.612 | | 1 | 1280 | 8192 | machete_bf16i4 | 0.026 | 0.810 | 203.219 | | 1 | 8192 | 1024 | bf16_baseline | 0.018 | 0.952 | 953.176 | | 1 | 8192 | 1024 | bf16_oss_fast_gemv | 0.015 | 1.100 | 1100.900 | | 1 | 8192 | 1024 | cuda_lite_fp8 | 0.014 | 1.198 | 600.054 | | 1 | 8192 | 1024 | marlin_bf16i4 | 0.015 | 1.144 | 287.150 | | 1 | 8192 | 1024 | machete_bf16i4 | 0.014 | 1.187 | 298.096 | | 1 | 7168 | 8192 | bf16_baseline | 0.073 | 1.609 | 1608.983 | | 1 | 7168 | 8192 | bf16_oss_fast_gemv | 0.069 | 1.697 | 1697.308 | | 1 | 7168 | 8192 | cuda_lite_fp8 | 0.044 | 2.679 | 1340.093 | | 1 | 7168 | 8192 | marlin_bf16i4 | 0.033 | 3.590 | 898.436 | | 1 | 7168 | 8192 | machete_bf16i4 | 0.039 | 3.017 | 755.147 | | 1 | 8192 | 3584 | bf16_baseline | 0.045 | 1.312 | 1312.239 | | 1 | 8192 | 3584 | bf16_oss_fast_gemv | 0.041 | 1.427 | 1427.166 | | 1 | 8192 | 3584 | cuda_lite_fp8 | 0.026 | 2.271 | 1136.151 | | 1 | 8192 | 3584 | marlin_bf16i4 | 0.021 | 2.808 | 703.164 | | 1 | 8192 | 3584 | machete_bf16i4 | 0.024 | 2.460 | 615.990 | Note that currently the precision with `fast_gemv` kernel and `cuda_lite` does not match yet. so need fp8 to coming in for a fairer result. Also no_cuda_graph flag enabled when running the quantize_bench heuristic sweep results from the 4 problem sizes we care about: P1722806148 **Next step:** Need fp8 mixed precision support for fast gemv kernel which is what we want Differential Revision: D68470488
YUNQIUGUO
force-pushed
the
export-D68470488
branch
from
d3aecb7 to
4cb9654
Compare
|
This pull request was exported from Phabricator. Differential Revision: D68470488 |
YUNQIUGUO
added a commit
to YUNQIUGUO/FBGEMM
that referenced
this pull request
Feb 4, 2025
Summary: X-link: facebookresearch/FBGEMM#688 This diff content includes: 1. Port OSS FastGEMV `bf16` kernel into fbcode and expose to python as a step 1 - `torch.ops.fbgemm.bf16_fast_gemv` https://github.com/wangsiping97/FastGEMV/blob/1fdff6f74aade033c02727a419afd6a4b4bfbc3f/fast_gemv.cu#L14 2. Add `bf16_oss_fast_gemv` to quantize ops benchmark script 3. Add two simple tests for custom op`torch.ops.fbgemm.f16_fast_gemv` to test - `torch.compile()` able - correctness Perf numbers compared with `bf16_baseline,bf16_oss_fast_gemv,cuda_lite,marlin_bf16i4,machete_bf16i4` ====================== ### Benchmark Results on H100 | **M** | **N** | **K** | **Method** | **Elapsed Time (ms)** | **TFLOPS** | **GB/s** | | --- | --- | --- | --- | --- | --- | --- | | 1 | 1280 | 8192 | bf16_baseline | 0.024 | 0.860 | 861.042 | | 1 | 1280 | 8192 | bf16_oss_fast_gemv | 0.019 | 1.126 | 1127.391 | | 1 | 1280 | 8192 | cuda_lite_fp8 | 0.015 | 1.357 | 679.032 | | 1 | 1280 | 8192 | marlin_bf16i4 | 0.027 | 0.768 | 192.612 | | 1 | 1280 | 8192 | machete_bf16i4 | 0.026 | 0.810 | 203.219 | | 1 | 8192 | 1024 | bf16_baseline | 0.018 | 0.952 | 953.176 | | 1 | 8192 | 1024 | bf16_oss_fast_gemv | 0.015 | 1.100 | 1100.900 | | 1 | 8192 | 1024 | cuda_lite_fp8 | 0.014 | 1.198 | 600.054 | | 1 | 8192 | 1024 | marlin_bf16i4 | 0.015 | 1.144 | 287.150 | | 1 | 8192 | 1024 | machete_bf16i4 | 0.014 | 1.187 | 298.096 | | 1 | 7168 | 8192 | bf16_baseline | 0.073 | 1.609 | 1608.983 | | 1 | 7168 | 8192 | bf16_oss_fast_gemv | 0.069 | 1.697 | 1697.308 | | 1 | 7168 | 8192 | cuda_lite_fp8 | 0.044 | 2.679 | 1340.093 | | 1 | 7168 | 8192 | marlin_bf16i4 | 0.033 | 3.590 | 898.436 | | 1 | 7168 | 8192 | machete_bf16i4 | 0.039 | 3.017 | 755.147 | | 1 | 8192 | 3584 | bf16_baseline | 0.045 | 1.312 | 1312.239 | | 1 | 8192 | 3584 | bf16_oss_fast_gemv | 0.041 | 1.427 | 1427.166 | | 1 | 8192 | 3584 | cuda_lite_fp8 | 0.026 | 2.271 | 1136.151 | | 1 | 8192 | 3584 | marlin_bf16i4 | 0.021 | 2.808 | 703.164 | | 1 | 8192 | 3584 | machete_bf16i4 | 0.024 | 2.460 | 615.990 | Note that currently the precision with `fast_gemv` kernel and `cuda_lite` does not match yet. so need fp8 to coming in for a fairer result. Also no_cuda_graph flag enabled when running the quantize_bench heuristic sweep results from the 4 problem sizes we care about: P1722806148 **Next step:** Need fp8 mixed precision support for fast gemv kernel which is what we want Differential Revision: D68470488
YUNQIUGUO
force-pushed
the
export-D68470488
branch
from
4cb9654 to
15549a3
Compare
|
This pull request was exported from Phabricator. Differential Revision: D68470488 |
Summary: X-link: facebookresearch/FBGEMM#688 This diff content includes: 1. Port OSS FastGEMV `bf16` kernel into fbcode and expose to python as a step 1 - `torch.ops.fbgemm.bf16_fast_gemv` https://github.com/wangsiping97/FastGEMV/blob/1fdff6f74aade033c02727a419afd6a4b4bfbc3f/fast_gemv.cu#L14 2. Add `bf16_oss_fast_gemv` to quantize ops benchmark script 3. Add two simple tests for custom op`torch.ops.fbgemm.f16_fast_gemv` to test - `torch.compile()` able - correctness Perf numbers compared with `bf16_baseline,bf16_oss_fast_gemv,cuda_lite,marlin_bf16i4,machete_bf16i4` ====================== ### Benchmark Results on H100 | **M** | **N** | **K** | **Method** | **Elapsed Time (ms)** | **TFLOPS** | **GB/s** | | --- | --- | --- | --- | --- | --- | --- | | 1 | 1280 | 8192 | bf16_baseline | 0.024 | 0.860 | 861.042 | | 1 | 1280 | 8192 | bf16_oss_fast_gemv | 0.019 | 1.126 | 1127.391 | | 1 | 1280 | 8192 | cuda_lite_fp8 | 0.015 | 1.357 | 679.032 | | 1 | 1280 | 8192 | marlin_bf16i4 | 0.027 | 0.768 | 192.612 | | 1 | 1280 | 8192 | machete_bf16i4 | 0.026 | 0.810 | 203.219 | | 1 | 8192 | 1024 | bf16_baseline | 0.018 | 0.952 | 953.176 | | 1 | 8192 | 1024 | bf16_oss_fast_gemv | 0.015 | 1.100 | 1100.900 | | 1 | 8192 | 1024 | cuda_lite_fp8 | 0.014 | 1.198 | 600.054 | | 1 | 8192 | 1024 | marlin_bf16i4 | 0.015 | 1.144 | 287.150 | | 1 | 8192 | 1024 | machete_bf16i4 | 0.014 | 1.187 | 298.096 | | 1 | 7168 | 8192 | bf16_baseline | 0.073 | 1.609 | 1608.983 | | 1 | 7168 | 8192 | bf16_oss_fast_gemv | 0.069 | 1.697 | 1697.308 | | 1 | 7168 | 8192 | cuda_lite_fp8 | 0.044 | 2.679 | 1340.093 | | 1 | 7168 | 8192 | marlin_bf16i4 | 0.033 | 3.590 | 898.436 | | 1 | 7168 | 8192 | machete_bf16i4 | 0.039 | 3.017 | 755.147 | | 1 | 8192 | 3584 | bf16_baseline | 0.045 | 1.312 | 1312.239 | | 1 | 8192 | 3584 | bf16_oss_fast_gemv | 0.041 | 1.427 | 1427.166 | | 1 | 8192 | 3584 | cuda_lite_fp8 | 0.026 | 2.271 | 1136.151 | | 1 | 8192 | 3584 | marlin_bf16i4 | 0.021 | 2.808 | 703.164 | | 1 | 8192 | 3584 | machete_bf16i4 | 0.024 | 2.460 | 615.990 | Note that currently the precision with `fast_gemv` kernel and `cuda_lite` does not match yet. so need fp8 to coming in for a fairer result. Also no_cuda_graph flag enabled when running the quantize_bench heuristic sweep results from the 4 problem sizes we care about: P1722806148 **Next step:** Need fp8 mixed precision support for fast gemv kernel which is what we want Differential Revision: D68470488
YUNQIUGUO
force-pushed
the
export-D68470488
branch
from
15549a3 to
0b15c28
Compare
|
This pull request was exported from Phabricator. Differential Revision: D68470488 |
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.
Summary:
This diff content includes:
fp16kernel into fbcode and expose to python as a step 1 -torch.ops.fbgemm.f16_fast_gemvhttps://github.com/wangsiping97/FastGEMV/blob/1fdff6f74aade033c02727a419afd6a4b4bfbc3f/fast_gemv.cu#L14
fp16_oss_fast_gemvto quantize ops benchmark scripttorch.ops.fbgemm.f16_fast_gemvto testtorch.compile()ableNext step:
Need fp8 mixed precision support for fast gemv kernel which is what we want
Differential Revision: D68470488