amd fp8 rowwise gemm prefill shape tuning #3607
Closed
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
✅ Deploy Preview for pytorch-fbgemm-docs ready!
To edit notification comments on pull requests, go to your Netlify site configuration. |
|
This pull request was exported from Phabricator. Differential Revision: D68521662 |
|
This pull request was exported from Phabricator. Differential Revision: D68521662 |
mxz297
added a commit
to mxz297/FBGEMM
that referenced
this pull request
Jan 23, 2025
Summary: X-link: facebookresearch/FBGEMM#685 Pull Request resolved: pytorch#3607 This diff aims to add a more robust FP8 rowwise heuristics for LLM, especially for prefill cases. Consider input [M, K] and weight [N, K]. For LLMs, N and K are fixed across different prefill/decode lengths. Thus the new heuristic is based on lookup for (N,K) and then do a range based lookup for M. For each combination of N and K, there is offline tuning for many M, looking like: ``` 5280, 8192, 3584, 0.318272, 256x256x192x128_16x16_8x6_intrawave_v3_kbatch_1 5312, 8192, 3584, 0.322179, 256x256x192x128_16x16_8x6_intrawave_v3_kbatch_1 5344, 8192, 3584, 0.320632, 256x256x192x128_16x16_8x6_intrawave_v3_kbatch_1 5376, 8192, 3584, 0.317728, 256x256x192x128_16x16_8x6_intrawave_v3_kbatch_1 5408, 8192, 3584, 0.338742, 256x224x256x128_16x16_7x8_intrawave_v3_kbatch_1 5440, 8192, 3584, 0.341432, 256x224x256x128_16x16_7x8_intrawave_v3_kbatch_1 5472, 8192, 3584, 0.3436, 256x224x256x128_16x16_7x8_intrawave_v3_kbatch_1 5536, 8192, 3584, 0.341703, 256x224x256x128_16x16_7x8_intrawave_v3_kbatch_1 5568, 8192, 3584, 0.342054, 256x224x256x128_16x16_7x8_intrawave_v3_kbatch_1 5632, 8192, 3584, 0.347904, 256x224x256x128_16x16_7x8_intrawave_v3_kbatch_1 5664, 8192, 3584, 0.345129, 256x224x256x128_16x16_7x8_intrawave_v3_kbatch_1 ``` A clear pattern is that a single instance is the top choice for a large range, justifying the M range based heuristic. The full tuning log is parsed and converted into a std::map for range based lookup. One key question here is which instance to use right at the range where the best instance has changed. For example: ``` 5376, 8192, 3584, 0.317728, 256x256x192x128_16x16_8x6_intrawave_v3_kbatch_1 5408, 8192, 3584, 0.338742, 256x224x256x128_16x16_7x8_intrawave_v3_kbatch_1 ``` Should we use 256x256x192x128 or 256x224x256x128 for M = 5377 to 5407? The implementation uses the tuning entry for the larger value (so use 256x224x256x128). The rational is if we use the smaller entry, it may lead to increased thread blocks and cause bad perf; in contrast, if we use the larger entry, the perf will in theory be the same as the larger entry itself. Empirically, using the smaller entry lead to some degraded perf for untuned values. Differential Revision: D68521662
mxz297
force-pushed
the
export-D68521662
branch
from
20e71d5 to
f4c6b0e
Compare
|
This pull request was exported from Phabricator. Differential Revision: D68521662 |
mxz297
added a commit
to mxz297/FBGEMM
that referenced
this pull request
Jan 24, 2025
Summary: X-link: facebookresearch/FBGEMM#685 Pull Request resolved: pytorch#3607 This diff aims to add a more robust FP8 rowwise heuristics for LLM, especially for prefill cases. Consider input [M, K] and weight [N, K]. For LLMs, N and K are fixed across different prefill/decode lengths. Thus the new heuristic is based on lookup for (N,K) and then do a range based lookup for M. For each combination of N and K, there is offline tuning for many M, looking like: ``` 5280, 8192, 3584, 0.318272, 256x256x192x128_16x16_8x6_intrawave_v3_kbatch_1 5312, 8192, 3584, 0.322179, 256x256x192x128_16x16_8x6_intrawave_v3_kbatch_1 5344, 8192, 3584, 0.320632, 256x256x192x128_16x16_8x6_intrawave_v3_kbatch_1 5376, 8192, 3584, 0.317728, 256x256x192x128_16x16_8x6_intrawave_v3_kbatch_1 5408, 8192, 3584, 0.338742, 256x224x256x128_16x16_7x8_intrawave_v3_kbatch_1 5440, 8192, 3584, 0.341432, 256x224x256x128_16x16_7x8_intrawave_v3_kbatch_1 5472, 8192, 3584, 0.3436, 256x224x256x128_16x16_7x8_intrawave_v3_kbatch_1 5536, 8192, 3584, 0.341703, 256x224x256x128_16x16_7x8_intrawave_v3_kbatch_1 5568, 8192, 3584, 0.342054, 256x224x256x128_16x16_7x8_intrawave_v3_kbatch_1 5632, 8192, 3584, 0.347904, 256x224x256x128_16x16_7x8_intrawave_v3_kbatch_1 5664, 8192, 3584, 0.345129, 256x224x256x128_16x16_7x8_intrawave_v3_kbatch_1 ``` A clear pattern is that a single instance is the top choice for a large range, justifying the M range based heuristic. The full tuning log is parsed and converted into a std::map for range based lookup. One key question here is which instance to use right at the range where the best instance has changed. For example: ``` 5376, 8192, 3584, 0.317728, 256x256x192x128_16x16_8x6_intrawave_v3_kbatch_1 5408, 8192, 3584, 0.338742, 256x224x256x128_16x16_7x8_intrawave_v3_kbatch_1 ``` Should we use 256x256x192x128 or 256x224x256x128 for M = 5377 to 5407? The implementation uses the tuning entry for the larger value (so use 256x224x256x128). The rational is if we use the smaller entry, it may lead to increased thread blocks and cause bad perf; in contrast, if we use the larger entry, the perf will in theory be the same as the larger entry itself. Empirically, using the smaller entry lead to some degraded perf for untuned values. Reviewed By: jwfromm Differential Revision: D68521662
mxz297
force-pushed
the
export-D68521662
branch
from
f4c6b0e to
9acfa9a
Compare
Summary: X-link: facebookresearch/FBGEMM#685 Pull Request resolved: pytorch#3607 This diff aims to add a more robust FP8 rowwise heuristics for LLM, especially for prefill cases. Consider input [M, K] and weight [N, K]. For LLMs, N and K are fixed across different prefill/decode lengths. Thus the new heuristic is based on lookup for (N,K) and then do a range based lookup for M. For each combination of N and K, there is offline tuning for many M, looking like: ``` 5280, 8192, 3584, 0.318272, 256x256x192x128_16x16_8x6_intrawave_v3_kbatch_1 5312, 8192, 3584, 0.322179, 256x256x192x128_16x16_8x6_intrawave_v3_kbatch_1 5344, 8192, 3584, 0.320632, 256x256x192x128_16x16_8x6_intrawave_v3_kbatch_1 5376, 8192, 3584, 0.317728, 256x256x192x128_16x16_8x6_intrawave_v3_kbatch_1 5408, 8192, 3584, 0.338742, 256x224x256x128_16x16_7x8_intrawave_v3_kbatch_1 5440, 8192, 3584, 0.341432, 256x224x256x128_16x16_7x8_intrawave_v3_kbatch_1 5472, 8192, 3584, 0.3436, 256x224x256x128_16x16_7x8_intrawave_v3_kbatch_1 5536, 8192, 3584, 0.341703, 256x224x256x128_16x16_7x8_intrawave_v3_kbatch_1 5568, 8192, 3584, 0.342054, 256x224x256x128_16x16_7x8_intrawave_v3_kbatch_1 5632, 8192, 3584, 0.347904, 256x224x256x128_16x16_7x8_intrawave_v3_kbatch_1 5664, 8192, 3584, 0.345129, 256x224x256x128_16x16_7x8_intrawave_v3_kbatch_1 ``` A clear pattern is that a single instance is the top choice for a large range, justifying the M range based heuristic. The full tuning log is parsed and converted into a std::map for range based lookup. One key question here is which instance to use right at the range where the best instance has changed. For example: ``` 5376, 8192, 3584, 0.317728, 256x256x192x128_16x16_8x6_intrawave_v3_kbatch_1 5408, 8192, 3584, 0.338742, 256x224x256x128_16x16_7x8_intrawave_v3_kbatch_1 ``` Should we use 256x256x192x128 or 256x224x256x128 for M = 5377 to 5407? The implementation uses the tuning entry for the larger value (so use 256x224x256x128). The rational is if we use the smaller entry, it may lead to increased thread blocks and cause bad perf; in contrast, if we use the larger entry, the perf will in theory be the same as the larger entry itself. Empirically, using the smaller entry lead to some degraded perf for untuned values. Reviewed By: jwfromm Differential Revision: D68521662
|
This pull request was exported from Phabricator. Differential Revision: D68521662 |
mxz297
force-pushed
the
export-D68521662
branch
from
9acfa9a to
9f84364
Compare
|
This pull request has been merged in 74490d6. |
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 aims to add a more robust FP8 rowwise heuristics for LLM, especially for prefill cases. Consider input [M, K] and weight [N, K].
For LLMs, N and K are fixed across different prefill/decode lengths. Thus the new heuristic is based on lookup for (N,K) and then do a range based lookup for M.
For each combination of N and K, there is offline tuning for many M, looking like:
A clear pattern is that a single instance is the top choice for a large range, justifying the M range based heuristic.
The full tuning log is parsed and converted into a std::map for range based lookup.
One key question here is which instance to use right at the range where the best instance has changed. For example:
Should we use 256x256x192x128 or 256x224x256x128 for M = 5377 to 5407? The implementation uses the tuning entry for the larger value (so use 256x224x256x128). The rational is if we use the smaller entry, it may lead to increased thread blocks and cause bad perf; in contrast, if we use the larger entry, the perf will in theory be the same as the larger entry itself. Empirically, using the smaller entry lead to some degraded perf for untuned values.
Differential Revision: D68521662