Prototype FP8 Blackwell persistent + TMA kernel with warp specialization #385
+207
−1
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
jananisriram
had a problem deploying
to
docker-s3-upload
GitHub Actions
Error
jananisriram
had a problem deploying
to
docker-s3-upload
GitHub Actions
Error
|
This pull request was exported from Phabricator. Differential Revision: D81470285 |
facebook-github-bot
force-pushed
the
export-D81470285
branch
from
e6578cb to
11ddc62
Compare
|
This pull request was exported from Phabricator. Differential Revision: D81470285 |
facebook-github-bot
had a problem deploying
to
docker-s3-upload
GitHub Actions
Failure
facebook-github-bot
had a problem deploying
to
docker-s3-upload
GitHub Actions
Failure
facebook-github-bot
pushed a commit
that referenced
this pull request
Sep 16, 2025
Summary: Taking inspiration from D77053488, Triton's [persistent matmul](https://triton-lang.org/main/getting-started/tutorials/09-persistent-matmul.html#) tutorial, and Triton's [block scaled matmul](https://triton-lang.org/main/getting-started/tutorials/10-block-scaled-matmul.html) tutorial, write and benchmark a persistent + TMA Triton kernel for FP8 workloads on Blackwell which enables warp specialization and flattening. Recall the following for FP8 workloads: - Input shapes: `torch.float8_e4m3fn` (`tl.float8e4nv`) - Output: `torch.float16` (per-tensor, `tl.float16`) or `torch.bfloat16` (per-row, `tl.bfloat16`) - Accumulation: `torch.float32` (`tl.float32`). Note the following limitations: - `(K, N)` (second matrix in GEMM) needs to be a multiple of 16 - `num_warps >= 4`: TMA instructions expect at least a 128-thread (4 warps) group. 1 warp = 32 threads, so we need 4 warps per thread block to ensure correct functionality. Note that the current kernel is being autotuned on just one config; this will be changed in a future diff. Differential Revision: D81470285
facebook-github-bot
force-pushed
the
export-D81470285
branch
from
11ddc62 to
8e64a26
Compare
facebook-github-bot
had a problem deploying
to
docker-s3-upload
GitHub Actions
Failure
facebook-github-bot
had a problem deploying
to
docker-s3-upload
GitHub Actions
Failure
|
@jananisriram has exported this pull request. If you are a Meta employee, you can view the originating diff in D81470285. |
facebook-github-bot
pushed a commit
that referenced
this pull request
Sep 16, 2025
Summary: Taking inspiration from D77053488, Triton's [persistent matmul](https://triton-lang.org/main/getting-started/tutorials/09-persistent-matmul.html#) tutorial, and Triton's [block scaled matmul](https://triton-lang.org/main/getting-started/tutorials/10-block-scaled-matmul.html) tutorial, write and benchmark a persistent + TMA Triton kernel for FP8 workloads on Blackwell which enables warp specialization and flattening. Recall the following for FP8 workloads: - Input shapes: `torch.float8_e4m3fn` (`tl.float8e4nv`) - Output: `torch.float16` (per-tensor, `tl.float16`) or `torch.bfloat16` (per-row, `tl.bfloat16`) - Accumulation: `torch.float32` (`tl.float32`). Note the following limitations: - `(K, N)` (second matrix in GEMM) needs to be a multiple of 16 - `num_warps >= 4`: TMA instructions expect at least a 128-thread (4 warps) group. 1 warp = 32 threads, so we need 4 warps per thread block to ensure correct functionality. Note that the current kernel is being autotuned on just one config; this will be changed in a future diff. Differential Revision: D81470285
facebook-github-bot
pushed a commit
that referenced
this pull request
Sep 16, 2025
Summary: Taking inspiration from D77053488, Triton's [persistent matmul](https://triton-lang.org/main/getting-started/tutorials/09-persistent-matmul.html#) tutorial, and Triton's [block scaled matmul](https://triton-lang.org/main/getting-started/tutorials/10-block-scaled-matmul.html) tutorial, write and benchmark a persistent + TMA Triton kernel for FP8 workloads on Blackwell which enables warp specialization and flattening. Recall the following for FP8 workloads: - Input shapes: `torch.float8_e4m3fn` (`tl.float8e4nv`) - Output: `torch.float16` (per-tensor, `tl.float16`) or `torch.bfloat16` (per-row, `tl.bfloat16`) - Accumulation: `torch.float32` (`tl.float32`). Note the following limitations: - `(K, N)` (second matrix in GEMM) needs to be a multiple of 16 - `num_warps >= 4`: TMA instructions expect at least a 128-thread (4 warps) group. 1 warp = 32 threads, so we need 4 warps per thread block to ensure correct functionality. Note that the current kernel is being autotuned on just one config; this will be changed in a future diff. Differential Revision: D81470285
jananisriram
added a commit
that referenced
this pull request
Sep 16, 2025
Summary: Taking inspiration from D77053488, Triton's [persistent matmul](https://triton-lang.org/main/getting-started/tutorials/09-persistent-matmul.html#) tutorial, and Triton's [block scaled matmul](https://triton-lang.org/main/getting-started/tutorials/10-block-scaled-matmul.html) tutorial, write and benchmark a persistent + TMA Triton kernel for FP8 workloads on Blackwell which enables warp specialization and flattening. Recall the following for FP8 workloads: - Input shapes: `torch.float8_e4m3fn` (`tl.float8e4nv`) - Output: `torch.float16` (per-tensor, `tl.float16`) or `torch.bfloat16` (per-row, `tl.bfloat16`) - Accumulation: `torch.float32` (`tl.float32`). Note the following limitations: - `(K, N)` (second matrix in GEMM) needs to be a multiple of 16 - `num_warps >= 4`: TMA instructions expect at least a 128-thread (4 warps) group. 1 warp = 32 threads, so we need 4 warps per thread block to ensure correct functionality. Note that the current kernel is being autotuned on just one config; this will be changed in a future diff. Differential Revision: D81470285
jananisriram
added a commit
that referenced
this pull request
Sep 16, 2025
Summary: Taking inspiration from D77053488, Triton's [persistent matmul](https://triton-lang.org/main/getting-started/tutorials/09-persistent-matmul.html#) tutorial, and Triton's [block scaled matmul](https://triton-lang.org/main/getting-started/tutorials/10-block-scaled-matmul.html) tutorial, write and benchmark a persistent + TMA Triton kernel for FP8 workloads on Blackwell which enables warp specialization and flattening. Recall the following for FP8 workloads: - Input shapes: `torch.float8_e4m3fn` (`tl.float8e4nv`) - Output: `torch.float16` (per-tensor, `tl.float16`) or `torch.bfloat16` (per-row, `tl.bfloat16`) - Accumulation: `torch.float32` (`tl.float32`). Note the following limitations: - `(K, N)` (second matrix in GEMM) needs to be a multiple of 16 - `num_warps >= 4`: TMA instructions expect at least a 128-thread (4 warps) group. 1 warp = 32 threads, so we need 4 warps per thread block to ensure correct functionality. Note that the current kernel is being autotuned on just one config; this will be changed in a future diff. Differential Revision: D81470285
jananisriram
force-pushed
the
export-D81470285
branch
from
8e64a26 to
13ceedb
Compare
jananisriram
had a problem deploying
to
docker-s3-upload
GitHub Actions
Failure
jananisriram
had a problem deploying
to
docker-s3-upload
GitHub Actions
Failure
|
@jananisriram has exported this pull request. If you are a Meta employee, you can view the originating diff in D81470285. |
jananisriram
added a commit
that referenced
this pull request
Sep 16, 2025
Summary: Taking inspiration from D77053488, Triton's [persistent matmul](https://triton-lang.org/main/getting-started/tutorials/09-persistent-matmul.html#) tutorial, and Triton's [block scaled matmul](https://triton-lang.org/main/getting-started/tutorials/10-block-scaled-matmul.html) tutorial, write and benchmark a persistent + TMA Triton kernel for FP8 workloads on Blackwell which enables warp specialization and flattening. Recall the following for FP8 workloads: - Input shapes: `torch.float8_e4m3fn` (`tl.float8e4nv`) - Output: `torch.float16` (per-tensor, `tl.float16`) or `torch.bfloat16` (per-row, `tl.bfloat16`) - Accumulation: `torch.float32` (`tl.float32`). Note the following limitations: - `(K, N)` (second matrix in GEMM) needs to be a multiple of 16 - `num_warps >= 4`: TMA instructions expect at least a 128-thread (4 warps) group. 1 warp = 32 threads, so we need 4 warps per thread block to ensure correct functionality. Note that the current kernel is being autotuned on just one config; this will be changed in a future diff. Differential Revision: D81470285
jananisriram
added a commit
that referenced
this pull request
Sep 16, 2025
Summary: Taking inspiration from D77053488, Triton's [persistent matmul](https://triton-lang.org/main/getting-started/tutorials/09-persistent-matmul.html#) tutorial, and Triton's [block scaled matmul](https://triton-lang.org/main/getting-started/tutorials/10-block-scaled-matmul.html) tutorial, write and benchmark a persistent + TMA Triton kernel for FP8 workloads on Blackwell which enables warp specialization and flattening. Recall the following for FP8 workloads: - Input shapes: `torch.float8_e4m3fn` (`tl.float8e4nv`) - Output: `torch.float16` (per-tensor, `tl.float16`) or `torch.bfloat16` (per-row, `tl.bfloat16`) - Accumulation: `torch.float32` (`tl.float32`). Note the following limitations: - `(K, N)` (second matrix in GEMM) needs to be a multiple of 16 - `num_warps >= 4`: TMA instructions expect at least a 128-thread (4 warps) group. 1 warp = 32 threads, so we need 4 warps per thread block to ensure correct functionality. Note that the current kernel is being autotuned on just one config; this will be changed in a future diff. Differential Revision: D81470285
jananisriram
added a commit
that referenced
this pull request
Sep 16, 2025
Summary: Taking inspiration from D77053488, Triton's [persistent matmul](https://triton-lang.org/main/getting-started/tutorials/09-persistent-matmul.html#) tutorial, and Triton's [block scaled matmul](https://triton-lang.org/main/getting-started/tutorials/10-block-scaled-matmul.html) tutorial, write and benchmark a persistent + TMA Triton kernel for FP8 workloads on Blackwell which enables warp specialization and flattening. Recall the following for FP8 workloads: - Input shapes: `torch.float8_e4m3fn` (`tl.float8e4nv`) - Output: `torch.float16` (per-tensor, `tl.float16`) or `torch.bfloat16` (per-row, `tl.bfloat16`) - Accumulation: `torch.float32` (`tl.float32`). Note the following limitations: - `(K, N)` (second matrix in GEMM) needs to be a multiple of 16 - `num_warps >= 4`: TMA instructions expect at least a 128-thread (4 warps) group. 1 warp = 32 threads, so we need 4 warps per thread block to ensure correct functionality. Note that the current kernel is being autotuned on just one config; this will be changed in a future diff. Differential Revision: D81470285
jananisriram
force-pushed
the
export-D81470285
branch
from
13ceedb to
ee54388
Compare
jananisriram
had a problem deploying
to
docker-s3-upload
GitHub Actions
Error
facebook-github-bot
had a problem deploying
to
docker-s3-upload
GitHub Actions
Failure
facebook-github-bot
had a problem deploying
to
docker-s3-upload
GitHub Actions
Failure
|
@jananisriram has exported this pull request. If you are a Meta employee, you can view the originating diff in D81470285. |
facebook-github-bot
pushed a commit
that referenced
this pull request
Sep 17, 2025
Summary: Taking inspiration from D77053488, Triton's [persistent matmul](https://triton-lang.org/main/getting-started/tutorials/09-persistent-matmul.html#) tutorial, and Triton's [block scaled matmul](https://triton-lang.org/main/getting-started/tutorials/10-block-scaled-matmul.html) tutorial, write and benchmark a persistent + TMA Triton kernel for FP8 workloads on Blackwell which enables warp specialization and flattening. Recall the following for FP8 workloads: - Input shapes: `torch.float8_e4m3fn` (`tl.float8e4nv`) - Output: `torch.float16` (per-tensor, `tl.float16`) or `torch.bfloat16` (per-row, `tl.bfloat16`) - Accumulation: `torch.float32` (`tl.float32`). Note the following limitations: - `(K, N)` (second matrix in GEMM) needs to be a multiple of 16 - `num_warps >= 4`: TMA instructions expect at least a 128-thread (4 warps) group. 1 warp = 32 threads, so we need 4 warps per thread block to ensure correct functionality. Note that the current kernel is being autotuned on just one config; this will be changed in a future diff. Differential Revision: D81470285
facebook-github-bot
pushed a commit
that referenced
this pull request
Sep 17, 2025
Summary: Taking inspiration from D77053488, Triton's [persistent matmul](https://triton-lang.org/main/getting-started/tutorials/09-persistent-matmul.html#) tutorial, and Triton's [block scaled matmul](https://triton-lang.org/main/getting-started/tutorials/10-block-scaled-matmul.html) tutorial, write and benchmark a persistent + TMA Triton kernel for FP8 workloads on Blackwell which enables warp specialization and flattening. Recall the following for FP8 workloads: - Input shapes: `torch.float8_e4m3fn` (`tl.float8e4nv`) - Output: `torch.float16` (per-tensor, `tl.float16`) or `torch.bfloat16` (per-row, `tl.bfloat16`) - Accumulation: `torch.float32` (`tl.float32`). Note the following limitations: - `(K, N)` (second matrix in GEMM) needs to be a multiple of 16 - `num_warps >= 4`: TMA instructions expect at least a 128-thread (4 warps) group. 1 warp = 32 threads, so we need 4 warps per thread block to ensure correct functionality. Note that the current kernel is being autotuned on just one config; this will be changed in a future diff. Differential Revision: D81470285
facebook-github-bot
force-pushed
the
export-D81470285
branch
from
0c68de7 to
aa5ffc7
Compare
facebook-github-bot
had a problem deploying
to
docker-s3-upload
GitHub Actions
Error
facebook-github-bot
had a problem deploying
to
docker-s3-upload
GitHub Actions
Error
|
@jananisriram has exported this pull request. If you are a Meta employee, you can view the originating diff in D81470285. |
facebook-github-bot
pushed a commit
that referenced
this pull request
Sep 17, 2025
Summary: Taking inspiration from D77053488, Triton's [persistent matmul](https://triton-lang.org/main/getting-started/tutorials/09-persistent-matmul.html#) tutorial, and Triton's [block scaled matmul](https://triton-lang.org/main/getting-started/tutorials/10-block-scaled-matmul.html) tutorial, write and benchmark a persistent + TMA Triton kernel for FP8 workloads on Blackwell which enables warp specialization and flattening. Recall the following for FP8 workloads: - Input shapes: `torch.float8_e4m3fn` (`tl.float8e4nv`) - Output: `torch.float16` (per-tensor, `tl.float16`) or `torch.bfloat16` (per-row, `tl.bfloat16`) - Accumulation: `torch.float32` (`tl.float32`). Note the following limitations: - `(K, N)` (second matrix in GEMM) needs to be a multiple of 16 - `num_warps >= 4`: TMA instructions expect at least a 128-thread (4 warps) group. 1 warp = 32 threads, so we need 4 warps per thread block to ensure correct functionality. Note that the current kernel is being autotuned on just one config; this will be changed in a future diff. Differential Revision: D81470285
facebook-github-bot
force-pushed
the
export-D81470285
branch
from
aa5ffc7 to
ef3c2a9
Compare
|
@jananisriram has exported this pull request. If you are a Meta employee, you can view the originating diff in D81470285. |
facebook-github-bot
had a problem deploying
to
docker-s3-upload
GitHub Actions
Failure
facebook-github-bot
had a problem deploying
to
docker-s3-upload
GitHub Actions
Failure
jananisriram
added a commit
that referenced
this pull request
Sep 17, 2025
Summary: Taking inspiration from D77053488, Triton's [persistent matmul](https://triton-lang.org/main/getting-started/tutorials/09-persistent-matmul.html#) tutorial, and Triton's [block scaled matmul](https://triton-lang.org/main/getting-started/tutorials/10-block-scaled-matmul.html) tutorial, write and benchmark a persistent + TMA Triton kernel for FP8 workloads on Blackwell which enables warp specialization and flattening. Recall the following for FP8 workloads: - Input shapes: `torch.float8_e4m3fn` (`tl.float8e4nv`) - Output: `torch.float16` (per-tensor, `tl.float16`) or `torch.bfloat16` (per-row, `tl.bfloat16`) - Accumulation: `torch.float32` (`tl.float32`). Note the following limitations: - `(K, N)` (second matrix in GEMM) needs to be a multiple of 16 - `num_warps >= 4`: TMA instructions expect at least a 128-thread (4 warps) group. 1 warp = 32 threads, so we need 4 warps per thread block to ensure correct functionality. Note that the current kernel is being autotuned on just one config; this will be changed in a future diff. Differential Revision: D81470285
jananisriram
added a commit
that referenced
this pull request
Sep 17, 2025
Summary: Taking inspiration from D77053488, Triton's [persistent matmul](https://triton-lang.org/main/getting-started/tutorials/09-persistent-matmul.html#) tutorial, and Triton's [block scaled matmul](https://triton-lang.org/main/getting-started/tutorials/10-block-scaled-matmul.html) tutorial, write and benchmark a persistent + TMA Triton kernel for FP8 workloads on Blackwell which enables warp specialization and flattening. Recall the following for FP8 workloads: - Input shapes: `torch.float8_e4m3fn` (`tl.float8e4nv`) - Output: `torch.float16` (per-tensor, `tl.float16`) or `torch.bfloat16` (per-row, `tl.bfloat16`) - Accumulation: `torch.float32` (`tl.float32`). Note the following limitations: - `(K, N)` (second matrix in GEMM) needs to be a multiple of 16 - `num_warps >= 4`: TMA instructions expect at least a 128-thread (4 warps) group. 1 warp = 32 threads, so we need 4 warps per thread block to ensure correct functionality. Note that the current kernel is being autotuned on just one config; this will be changed in a future diff. Differential Revision: D81470285
Summary: Taking inspiration from D77053488, Triton's [persistent matmul](https://triton-lang.org/main/getting-started/tutorials/09-persistent-matmul.html#) tutorial, and Triton's [block scaled matmul](https://triton-lang.org/main/getting-started/tutorials/10-block-scaled-matmul.html) tutorial, write and benchmark a persistent + TMA Triton kernel for FP8 workloads on Blackwell which enables warp specialization and flattening. Recall the following for FP8 workloads: - Input shapes: `torch.float8_e4m3fn` (`tl.float8e4nv`) - Output: `torch.float16` (per-tensor, `tl.float16`) or `torch.bfloat16` (per-row, `tl.bfloat16`) - Accumulation: `torch.float32` (`tl.float32`). Note the following limitations: - `(K, N)` (second matrix in GEMM) needs to be a multiple of 16 - `num_warps >= 4`: TMA instructions expect at least a 128-thread (4 warps) group. 1 warp = 32 threads, so we need 4 warps per thread block to ensure correct functionality. Note that the current kernel is being autotuned on just one config; this will be changed in a future diff. Reviewed By: njriasan Differential Revision: D81470285
facebook-github-bot
force-pushed
the
export-D81470285
branch
from
ef3c2a9 to
161229f
Compare
facebook-github-bot
temporarily deployed
to
docker-s3-upload
GitHub Actions
Inactive
facebook-github-bot
temporarily deployed
to
docker-s3-upload
GitHub Actions
Inactive
|
@jananisriram has exported this pull request. If you are a Meta employee, you can view the originating diff in D81470285. |
jananisriram
added a commit
that referenced
this pull request
Sep 18, 2025
Summary: Taking inspiration from D77053488, Triton's [persistent matmul](https://triton-lang.org/main/getting-started/tutorials/09-persistent-matmul.html#) tutorial, and Triton's [block scaled matmul](https://triton-lang.org/main/getting-started/tutorials/10-block-scaled-matmul.html) tutorial, write and benchmark a persistent + TMA Triton kernel for FP8 workloads on Blackwell which enables warp specialization and flattening. Recall the following for FP8 workloads: - Input shapes: `torch.float8_e4m3fn` (`tl.float8e4nv`) - Output: `torch.float16` (per-tensor, `tl.float16`) or `torch.bfloat16` (per-row, `tl.bfloat16`) - Accumulation: `torch.float32` (`tl.float32`). Note the following limitations: - `(K, N)` (second matrix in GEMM) needs to be a multiple of 16 - `num_warps >= 4`: TMA instructions expect at least a 128-thread (4 warps) group. 1 warp = 32 threads, so we need 4 warps per thread block to ensure correct functionality. Note that the current kernel is being autotuned on just one config; this will be changed in a future diff. Reviewed By: njriasan Differential Revision: D81470285
facebook-github-bot
pushed a commit
that referenced
this pull request
Sep 18, 2025
Summary: Taking inspiration from D77053488, Triton's [persistent matmul](https://triton-lang.org/main/getting-started/tutorials/09-persistent-matmul.html#) tutorial, and Triton's [block scaled matmul](https://triton-lang.org/main/getting-started/tutorials/10-block-scaled-matmul.html) tutorial, write and benchmark a persistent + TMA Triton kernel for FP8 workloads on Blackwell which enables warp specialization and flattening. Recall the following for FP8 workloads: - Input shapes: `torch.float8_e4m3fn` (`tl.float8e4nv`) - Output: `torch.float16` (per-tensor, `tl.float16`) or `torch.bfloat16` (per-row, `tl.bfloat16`) - Accumulation: `torch.float32` (`tl.float32`). Note the following limitations: - `(K, N)` (second matrix in GEMM) needs to be a multiple of 16 - `num_warps >= 4`: TMA instructions expect at least a 128-thread (4 warps) group. 1 warp = 32 threads, so we need 4 warps per thread block to ensure correct functionality. Note that the current kernel is being autotuned on just one config; this will be changed in a future diff. Reviewed By: njriasan Differential Revision: D81470285
facebook-github-bot
pushed a commit
that referenced
this pull request
Sep 18, 2025
Summary: Taking inspiration from D77053488, Triton's [persistent matmul](https://triton-lang.org/main/getting-started/tutorials/09-persistent-matmul.html#) tutorial, and Triton's [block scaled matmul](https://triton-lang.org/main/getting-started/tutorials/10-block-scaled-matmul.html) tutorial, write and benchmark a persistent + TMA Triton kernel for FP8 workloads on Blackwell which enables warp specialization and flattening. Recall the following for FP8 workloads: - Input shapes: `torch.float8_e4m3fn` (`tl.float8e4nv`) - Output: `torch.float16` (per-tensor, `tl.float16`) or `torch.bfloat16` (per-row, `tl.bfloat16`) - Accumulation: `torch.float32` (`tl.float32`). Note the following limitations: - `(K, N)` (second matrix in GEMM) needs to be a multiple of 16 - `num_warps >= 4`: TMA instructions expect at least a 128-thread (4 warps) group. 1 warp = 32 threads, so we need 4 warps per thread block to ensure correct functionality. Note that the current kernel is being autotuned on just one config; this will be changed in a future diff. Reviewed By: njriasan Differential Revision: D81470285
NikhilAPatel
approved these changes
Sep 18, 2025
|
@pytorchbot merge |
|
Mergebot is not configured for this repository. Please use the merge button provided by GitHub. |
|
@pytorchbot merge |
|
Mergebot is not configured for this repository. Please use the merge button provided by GitHub. |
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:
Taking inspiration from D77053488, Triton's persistent matmul tutorial, and Triton's block scaled matmul tutorial, write and benchmark a persistent + TMA Triton kernel for FP8 workloads on Blackwell which enables warp specialization and flattening.
Note the following limitations:
(K, N)(second matrix in GEMM) needs to be a multiple of 16num_warps >= 4: TMA instructions expect at least a 128-thread (4 warps) group. 1 warp = 32 threads, so we need 4 warps per thread block to ensure correct functionality.Differential Revision: D81470285