Kernels:
Trusted publisher
Uploaded using `kernel-builder`.
Browse files- build/torch-stable-abi29-cu126-x86_64-linux/{_flash_attn3_cuda_90fb2c3.abi3.so → _flash_attn3_cuda_fe2bb66.abi3.so} +2 -2
- build/torch-stable-abi29-cu126-x86_64-linux/_ops.py +3 -3
- build/torch-stable-abi29-cu126-x86_64-linux/metadata.json +8 -7
- build/torch-stable-abi29-cu126-x86_64-linux/metadata.json.sigstore +1 -1
- build/torch-stable-abi29-cu126-x86_64-linux/symbols.json +745 -0
- build/torch-stable-abi29-cu128-x86_64-linux/__init__.py +20 -0
- build/{torch-stable-abi29-cu130-x86_64-linux/_flash_attn3_cuda_90fb2c3.abi3.so → torch-stable-abi29-cu128-x86_64-linux/_flash_attn3_cuda_fe2bb66.abi3.so} +2 -2
- build/torch-stable-abi29-cu128-x86_64-linux/_ops.py +9 -0
- build/torch-stable-abi29-cu128-x86_64-linux/flash_attn_config.py +7 -0
- build/torch-stable-abi29-cu128-x86_64-linux/flash_attn_interface.py +1127 -0
- build/torch-stable-abi29-cu128-x86_64-linux/metadata.json +38 -0
- build/torch-stable-abi29-cu128-x86_64-linux/metadata.json.sigstore +1 -0
- build/torch-stable-abi29-cu128-x86_64-linux/symbols.json +745 -0
- build/torch-stable-abi29-cu130-x86_64-linux/_flash_attn3_cuda_fe2bb66.abi3.so +3 -0
- build/torch-stable-abi29-cu130-x86_64-linux/_ops.py +3 -3
- build/torch-stable-abi29-cu130-x86_64-linux/metadata.json +8 -7
- build/torch-stable-abi29-cu130-x86_64-linux/metadata.json.sigstore +1 -1
- build/torch-stable-abi29-cu130-x86_64-linux/symbols.json +745 -0
build/torch-stable-abi29-cu126-x86_64-linux/{_flash_attn3_cuda_90fb2c3.abi3.so → _flash_attn3_cuda_fe2bb66.abi3.so}
RENAMED
|
@@ -1,3 +1,3 @@
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
-
oid sha256:
|
| 3 |
-
size
|
|
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:039d4f5728ab690a142e92f1437c10aeacabe4f4a86cd942cf5b4aaaefe7c30d
|
| 3 |
+
size 526430928
|
build/torch-stable-abi29-cu126-x86_64-linux/_ops.py
CHANGED
|
@@ -1,9 +1,9 @@
|
|
| 1 |
import torch
|
| 2 |
-
from . import
|
| 3 |
-
ops = torch.ops.
|
| 4 |
|
| 5 |
def add_op_namespace_prefix(op_name: str):
|
| 6 |
"""
|
| 7 |
Prefix op by namespace.
|
| 8 |
"""
|
| 9 |
-
return f"
|
|
|
|
| 1 |
import torch
|
| 2 |
+
from . import _flash_attn3_cuda_fe2bb66
|
| 3 |
+
ops = torch.ops._flash_attn3_cuda_fe2bb66
|
| 4 |
|
| 5 |
def add_op_namespace_prefix(op_name: str):
|
| 6 |
"""
|
| 7 |
Prefix op by namespace.
|
| 8 |
"""
|
| 9 |
+
return f"_flash_attn3_cuda_fe2bb66::{op_name}"
|
build/torch-stable-abi29-cu126-x86_64-linux/metadata.json
CHANGED
|
@@ -1,6 +1,6 @@
|
|
| 1 |
{
|
| 2 |
"name": "flash-attn3",
|
| 3 |
-
"id": "
|
| 4 |
"version": 2,
|
| 5 |
"kernels-minver": "0.14.0",
|
| 6 |
"license": "BSD-3-Clause",
|
|
@@ -17,20 +17,21 @@
|
|
| 17 |
"algorithm": "sha256",
|
| 18 |
"files": {
|
| 19 |
"__init__.py": "b/45TnpaDDuTwp84ILd3Q+pQCIk2RnHtSSZKRq9Lj5E=",
|
| 20 |
-
"
|
| 21 |
-
"_ops.py": "/
|
| 22 |
"flash_attn_config.py": "uxo+eyMcDit//8YDaS4Rtk5BTMKzTSK9tOKTjdWVe/w=",
|
| 23 |
-
"flash_attn_interface.py": "y0bD2JYGMFgX6iDre3zJN2sBPuq/12N6Mo55Z+HQCNQ="
|
|
|
|
| 24 |
}
|
| 25 |
},
|
| 26 |
"provenance": {
|
| 27 |
"kernel-builder": {
|
| 28 |
-
"version": "0.
|
| 29 |
-
"commit": "
|
| 30 |
"dirty": false
|
| 31 |
},
|
| 32 |
"kernel": {
|
| 33 |
-
"commit": "
|
| 34 |
"dirty": false
|
| 35 |
}
|
| 36 |
}
|
|
|
|
| 1 |
{
|
| 2 |
"name": "flash-attn3",
|
| 3 |
+
"id": "_flash_attn3_cuda_fe2bb66",
|
| 4 |
"version": 2,
|
| 5 |
"kernels-minver": "0.14.0",
|
| 6 |
"license": "BSD-3-Clause",
|
|
|
|
| 17 |
"algorithm": "sha256",
|
| 18 |
"files": {
|
| 19 |
"__init__.py": "b/45TnpaDDuTwp84ILd3Q+pQCIk2RnHtSSZKRq9Lj5E=",
|
| 20 |
+
"_flash_attn3_cuda_fe2bb66.abi3.so": "A51PVyiraQoULpLxQ3wQrqyr5PSobNlCz1tKqu/nww0=",
|
| 21 |
+
"_ops.py": "qE76YSn3lYonVMHeyk13xZljOBdqQRapvPt15D/lvSI=",
|
| 22 |
"flash_attn_config.py": "uxo+eyMcDit//8YDaS4Rtk5BTMKzTSK9tOKTjdWVe/w=",
|
| 23 |
+
"flash_attn_interface.py": "y0bD2JYGMFgX6iDre3zJN2sBPuq/12N6Mo55Z+HQCNQ=",
|
| 24 |
+
"symbols.json": "dGYI9GazLj56q85OhdJTeGBnq6b9RzaJ911NQsrsjhs="
|
| 25 |
}
|
| 26 |
},
|
| 27 |
"provenance": {
|
| 28 |
"kernel-builder": {
|
| 29 |
+
"version": "0.18.0-dev0",
|
| 30 |
+
"commit": "d054906a13710b811b96697b4635f03020842733",
|
| 31 |
"dirty": false
|
| 32 |
},
|
| 33 |
"kernel": {
|
| 34 |
+
"commit": "fe2bb66fee6c2a77eccf08a0508dc57428d488ed",
|
| 35 |
"dirty": false
|
| 36 |
}
|
| 37 |
}
|
build/torch-stable-abi29-cu126-x86_64-linux/metadata.json.sigstore
CHANGED
|
@@ -1 +1 @@
|
|
| 1 |
-
{"mediaType":"application/vnd.dev.sigstore.bundle.v0.3+json", "verificationMaterial":{"certificate":{"rawBytes":"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"}, "tlogEntries":[{"logIndex":"3036565843", "logId":{"keyId":"wNI9atQGlz+VWfO6LRygH4QUfY/8W4RFwiT5i5WRgB0="}, "kindVersion":{"kind":"hashedrekord", "version":"0.0.1"}, "integratedTime":"1790868314", "inclusionPromise":{"signedEntryTimestamp":"MEUCIEZQLFgC3O9n8RCyLCNOCpO1DLzkiM74WlHkHT9kJCVlAiEA4qUqZBwBvVM8EFZ/rS18X0bhj5Ky/AVwoDO8y3qBfNo="}, "inclusionProof":{"logIndex":"2914661581", "rootHash":"44Gra7idm5DaZAoYVOUsrMsVEJpspWTqAvMHKzIC/Ac=", "treeSize":"2914661878", "hashes":["vZLqKVSgKN9j8f4d8msZUhG1U+UKBkfqfko/wQ1g9NY=", "JNNXvh7OgMPoGWz8NiPKfyqcCauAHP6H1OXaYwzwoBg=", "lEHO696XwKKismMI+Ww1ZbMbLvmaHbPx+K4zU4PEObo=", "DcBffV8c1qjistR2ax0aRRj+yPMuyemF8EFqGEj2IIE=", "bizQZwWkXZUbFdKw2+3WoKBczZ8Iud5lXP341bSZCMg=", "udnFf8M9gPVTBiBLm6SmeaAp0mkOa6bHGK/xCGE7q9Q=", "SHq+YRnKA2zeuthZNLzuxR5szADCiPL5gFlEnPsp0s0=", "A17aHema58n+waKcyKPjHICQw4lUkrke3RJ2+X43jXU=", "AUtbjOQuFHtYX0eoNN5HWh55kt6hgEc9SG9t8O3DSjs=", "WXt1IcXPA49H96t9Z3ItfyqY4Z6slXsSWdqhfhpKl+Y=", "lTLoVRjXSYVbDKJgTo23loiYc3RqJdtv1AgmoVyDKlc=", "7Tzd3tvRwcVZ82zfO1MExclwm0m1ZsS2AThT/K2Tlhg=", "x+oKrxjE+hawZze3Rc5uQ+ELyYoPBRbq8pWnKfLxvPw=", "EkYZ1Q2C2GzLyxAWt8wicxm9bAs4SIiWCFwJIkXBGBw=", "W4cMs02ycm0mIMwybrVXQENsvdY9LCPKwWAnSZsKd9E=", "Xu3SV3sfB3Jk5ygvrtnnDCdDb068HxGQBt+eg1oZ+Ts=", "hjmVlWMHU8ThyhWdZrPdgN2SCOrL/5LOxhgwDX0GFr4=", "cIz1OY2tO8cVEgp0TmAffPApps2hJuAcAN+pESrpHHY=", "4lYk7cX8PszZLoyQMWjMDB1BOBn4f7hehtFiHo92e5Y=", "ttpxd0TJprOUtb/8Wf5DeIzwNtUw72Q0YnXE2LFMIZo=", "qxzHanAzz57SDdmJe0B7bJK72NTIBbwEMGBBKvDOROw=", "xH/DCseLHr9eKoYT8qsORZK7zVdEGYWHuVtsVrD95wY="], "checkpoint":{"envelope":"rekor.sigstore.dev - 1193050959916656506\n2914661878\n44Gra7idm5DaZAoYVOUsrMsVEJpspWTqAvMHKzIC/Ac=\n\n— rekor.sigstore.dev wNI9ajBFAiBZNkygwxg6eyZxKjFTkM1sF5XHwwRHy0mnSPx6Vn8b6wIhAPOhaE9SCTDzSfLEyo1DrJM4RJIrRenxg7bOz1x5EMAo\n"}}, "canonicalizedBody":"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"}], "timestampVerificationData":{"rfc3161Timestamps":[{"signedTimestamp":"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"}]}}, "messageSignature":{"messageDigest":{"algorithm":"SHA2_256", "digest":"3oV9/+j/EOTS8AyQlzupSa8Ej1ux8JllRwiKrwL2fVg="}, "signature":"MEUCIHygQZFOptGH1o2D+Al24iAWnGo+JD8MUEREIaUx15vRAiEA6ocswxL/OnQy+mzGXedf8hwKfxMY5n2+G4HZ9WHlPGg="}}
|
|
|
|
| 1 |
+
{"mediaType":"application/vnd.dev.sigstore.bundle.v0.3+json", "verificationMaterial":{"certificate":{"rawBytes":"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"}, "tlogEntries":[{"logIndex":"3322704470", "logId":{"keyId":"wNI9atQGlz+VWfO6LRygH4QUfY/8W4RFwiT5i5WRgB0="}, "kindVersion":{"kind":"hashedrekord", "version":"0.0.1"}, "integratedTime":"1792423924", "inclusionPromise":{"signedEntryTimestamp":"MEUCIQDTXA4wRwF+Da7od1UF8uouZHAwNndBy+5w0MWAqPxw9QIgGOesr0KkhQitkPkbVDbRaVNV/5oTBjC1dRwjvK7/kbE="}, "inclusionProof":{"logIndex":"3200800208", "rootHash":"JIBrvAkPCOdONyvdEPWLiJaVhgOenfCrVu+BR/LcE5o=", "treeSize":"3200800470", "hashes":["yNYjEejhKPpJJ6e8flHVQFc5eTPPkHROz6tVh7WJFwQ=", "ei80ruJG4+GAnroeaRfXmcElivtIaObvRWutNh7xSaM=", "pVcQFgvRbgbgh+30tpsnUfusaScAJa8P4Yuz9Z9rcdw=", "s5fIu0/bVN6kxkBKNyQ5ccmJSCDOWkaihqtFaFl5JbQ=", "tRiVKYwgbrwnaC15xz6gcSvIo8X08M2nApdUBKIXltc=", "hlfl8q/LBBA/yPKRP80oKjym4PN7+RGq5RSxvyipGVw=", "wMHvC5wpk2w2QV75D6Eybt7sPv5Aocz4+mayMtfMwrw=", "HDkXM99Ho8nDP51E4bOExXg28DTNNAhQ/La81G1C/Tw=", "R8TlNGwBihyl19giHKZLJEREugagU2rNAydvQMl2gBo=", "iA4kmMobK6kdDo165OBDUjf233zk1esdXkmcmR1rF2I=", "Y/kTXkjjF16KI/qmcw9vLu8ZBoFtGZRa17fGXCGRAp0=", "11lr1PAKx8EF0MQ8eyTG2Okak7VNiBdR4LabBugomSw=", "uGEIH4dmv632MUSdl5AVDq3lFz9EPX5Y4BnnOA6WyoM=", "D37E7MwfvuENRJ4BUsqOTbzKKFjz+Dkh4kKr7xMQyU8=", "7WFpoaiQCzLwM3o2YI5Ngo8J7JD0OfyCCQtmXhWxVes=", "nyQV/+cWQUG5Hb2PKjQQNkQIJmz2qzAZKBQlkHGIknM=", "G+nk7mgE5h6dUmv2UY7NDJrY3WvrZY31raaO1AETt+o=", "qlTWNDzERnTBQVcHHXXVPjiUP0SsikGxfcxVlgqVFDo=", "gg9GNApNke3+zNhl4tsWtmIycCFrOMCB3lys6XuSQNQ=", "NlOclldmKPA1tMKIskqHAwuQDYHp31YwpK+OOMrMccc=", "qxzHanAzz57SDdmJe0B7bJK72NTIBbwEMGBBKvDOROw=", "xH/DCseLHr9eKoYT8qsORZK7zVdEGYWHuVtsVrD95wY="], "checkpoint":{"envelope":"rekor.sigstore.dev - 1193050959916656506\n3200800470\nJIBrvAkPCOdONyvdEPWLiJaVhgOenfCrVu+BR/LcE5o=\n\n— rekor.sigstore.dev wNI9ajBFAiAYYK66EewT2aqU/DVi3saKeJrsaQDWiLalNdvlOdUSPQIhAIiF9zj0f34NSQs7xJiJkhQezOtI31s3VCuNwR5TPJ15\n"}}, "canonicalizedBody":"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"}], "timestampVerificationData":{"rfc3161Timestamps":[{"signedTimestamp":"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"}]}}, "messageSignature":{"messageDigest":{"algorithm":"SHA2_256", "digest":"awT6AJo8oeQSvlgkBad6S3qUgx7OzaXN/D2FzrQx9xc="}, "signature":"MEYCIQDjRF0GgRL0nysnOmXJAB3QM78bdmyrbZVwl84ccJQSPAIhALA7TNs5lpLRnAQ8vmMUuGY9zlj+YSUJ6fPAtDqBBkN1"}}
|
build/torch-stable-abi29-cu126-x86_64-linux/symbols.json
ADDED
|
@@ -0,0 +1,745 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"schema_version": 1,
|
| 3 |
+
"module": "_flash_attn3_cuda_fe2bb66",
|
| 4 |
+
"functions": [
|
| 5 |
+
{
|
| 6 |
+
"name": "add_op_namespace_prefix",
|
| 7 |
+
"qualname": "add_op_namespace_prefix",
|
| 8 |
+
"module": "_flash_attn3_cuda_fe2bb66._ops",
|
| 9 |
+
"kind": "function",
|
| 10 |
+
"doc": "Prefix op by namespace.",
|
| 11 |
+
"signature": {
|
| 12 |
+
"parameters": [
|
| 13 |
+
{
|
| 14 |
+
"name": "op_name",
|
| 15 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 16 |
+
"default": null,
|
| 17 |
+
"annotation": "str"
|
| 18 |
+
}
|
| 19 |
+
],
|
| 20 |
+
"return_annotation": null
|
| 21 |
+
}
|
| 22 |
+
},
|
| 23 |
+
{
|
| 24 |
+
"name": "flash_attn_combine",
|
| 25 |
+
"qualname": "flash_attn_combine",
|
| 26 |
+
"module": "_flash_attn3_cuda_fe2bb66.flash_attn_interface",
|
| 27 |
+
"kind": "function",
|
| 28 |
+
"doc": null,
|
| 29 |
+
"signature": {
|
| 30 |
+
"parameters": [
|
| 31 |
+
{
|
| 32 |
+
"name": "out_partial",
|
| 33 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 34 |
+
"default": null,
|
| 35 |
+
"annotation": null
|
| 36 |
+
},
|
| 37 |
+
{
|
| 38 |
+
"name": "lse_partial",
|
| 39 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 40 |
+
"default": null,
|
| 41 |
+
"annotation": null
|
| 42 |
+
},
|
| 43 |
+
{
|
| 44 |
+
"name": "out",
|
| 45 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 46 |
+
"default": "None",
|
| 47 |
+
"annotation": null
|
| 48 |
+
},
|
| 49 |
+
{
|
| 50 |
+
"name": "out_dtype",
|
| 51 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 52 |
+
"default": "None",
|
| 53 |
+
"annotation": null
|
| 54 |
+
}
|
| 55 |
+
],
|
| 56 |
+
"return_annotation": null
|
| 57 |
+
}
|
| 58 |
+
},
|
| 59 |
+
{
|
| 60 |
+
"name": "flash_attn_func",
|
| 61 |
+
"qualname": "flash_attn_func",
|
| 62 |
+
"module": "_flash_attn3_cuda_fe2bb66.flash_attn_interface",
|
| 63 |
+
"kind": "function",
|
| 64 |
+
"doc": "dropout_p should be set to 0.0 during evaluation\nSupports multi-query and grouped-query attention (MQA/GQA) by passing in KV with fewer heads\nthan Q. Note that the number of heads in Q must be divisible by the number of heads in KV.\nFor example, if Q has 6 heads and K, V have 2 heads, head 0, 1, 2 of Q will attention to head\n0 of K, V, and head 3, 4, 5 of Q will attention to head 1 of K, V.\n\nIf causal=True, the causal mask is aligned to the bottom right corner of the attention matrix.\nFor example, if seqlen_q = 2 and seqlen_k = 5, the causal mask (1 = keep, 0 = masked out) is:\n 1 1 1 1 0\n 1 1 1 1 1\nIf seqlen_q = 5 and seqlen_k = 2, the causal mask is:\n 0 0\n 0 0\n 0 0\n 1 0\n 1 1\nIf the row of the mask is all zero, the output will be zero.\n\nIf window_size != (-1, -1), implements sliding window local attention. Query at position i\nwill only attend to keys between\n[i + seqlen_k - seqlen_q - window_size[0], i + seqlen_k - seqlen_q + window_size[1]] inclusive.\n\nArguments:\n q: (batch_size, seqlen, nheads, headdim)\n k: (batch_size, seqlen, nheads_k, headdim)\n v: (batch_size, seqlen, nheads_k, headdim)\n dropout_p: float. Dropout probability.\n softmax_scale: float. The scaling of QK^T before applying softmax.\n Default to 1 / sqrt(headdim).\n causal: bool. Whether to apply causal attention mask (e.g., for auto-regressive modeling).\n window_size: (left, right). If not (-1, -1), implements sliding window local attention.\n alibi_slopes: (nheads,) or (batch_size, nheads), fp32. A bias of\n (-alibi_slope * |i + seqlen_k - seqlen_q - j|)\n is added to the attention score of query i and key j.\n deterministic: bool. Whether to use the deterministic implementation of the backward pass,\n which is slightly slower and uses more memory. The forward pass is always deterministic.\n return_attn_probs: bool. Whether to return the attention probabilities. This option is for\n testing only. The returned probabilities are not guaranteed to be correct\n (they might not have the right scaling).\nReturn:\n out: (batch_size, seqlen, nheads, headdim).\n softmax_lse [optional, if return_attn_probs=True]: (batch_size, nheads, seqlen). The\n logsumexp of each row of the matrix QK^T * scaling (e.g., log of the softmax\n normalization factor).",
|
| 65 |
+
"signature": {
|
| 66 |
+
"parameters": [
|
| 67 |
+
{
|
| 68 |
+
"name": "q",
|
| 69 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 70 |
+
"default": null,
|
| 71 |
+
"annotation": null
|
| 72 |
+
},
|
| 73 |
+
{
|
| 74 |
+
"name": "k",
|
| 75 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 76 |
+
"default": null,
|
| 77 |
+
"annotation": null
|
| 78 |
+
},
|
| 79 |
+
{
|
| 80 |
+
"name": "v",
|
| 81 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 82 |
+
"default": null,
|
| 83 |
+
"annotation": null
|
| 84 |
+
},
|
| 85 |
+
{
|
| 86 |
+
"name": "softmax_scale",
|
| 87 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 88 |
+
"default": "None",
|
| 89 |
+
"annotation": null
|
| 90 |
+
},
|
| 91 |
+
{
|
| 92 |
+
"name": "causal",
|
| 93 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 94 |
+
"default": "False",
|
| 95 |
+
"annotation": null
|
| 96 |
+
},
|
| 97 |
+
{
|
| 98 |
+
"name": "qv",
|
| 99 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 100 |
+
"default": "None",
|
| 101 |
+
"annotation": null
|
| 102 |
+
},
|
| 103 |
+
{
|
| 104 |
+
"name": "q_descale",
|
| 105 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 106 |
+
"default": "None",
|
| 107 |
+
"annotation": null
|
| 108 |
+
},
|
| 109 |
+
{
|
| 110 |
+
"name": "k_descale",
|
| 111 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 112 |
+
"default": "None",
|
| 113 |
+
"annotation": null
|
| 114 |
+
},
|
| 115 |
+
{
|
| 116 |
+
"name": "v_descale",
|
| 117 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 118 |
+
"default": "None",
|
| 119 |
+
"annotation": null
|
| 120 |
+
},
|
| 121 |
+
{
|
| 122 |
+
"name": "window_size",
|
| 123 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 124 |
+
"default": "(-1, -1)",
|
| 125 |
+
"annotation": null
|
| 126 |
+
},
|
| 127 |
+
{
|
| 128 |
+
"name": "attention_chunk",
|
| 129 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 130 |
+
"default": "0",
|
| 131 |
+
"annotation": null
|
| 132 |
+
},
|
| 133 |
+
{
|
| 134 |
+
"name": "softcap",
|
| 135 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 136 |
+
"default": "0.0",
|
| 137 |
+
"annotation": null
|
| 138 |
+
},
|
| 139 |
+
{
|
| 140 |
+
"name": "num_splits",
|
| 141 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 142 |
+
"default": "1",
|
| 143 |
+
"annotation": null
|
| 144 |
+
},
|
| 145 |
+
{
|
| 146 |
+
"name": "pack_gqa",
|
| 147 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 148 |
+
"default": "None",
|
| 149 |
+
"annotation": null
|
| 150 |
+
},
|
| 151 |
+
{
|
| 152 |
+
"name": "deterministic",
|
| 153 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 154 |
+
"default": "False",
|
| 155 |
+
"annotation": null
|
| 156 |
+
},
|
| 157 |
+
{
|
| 158 |
+
"name": "sm_margin",
|
| 159 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 160 |
+
"default": "0",
|
| 161 |
+
"annotation": null
|
| 162 |
+
},
|
| 163 |
+
{
|
| 164 |
+
"name": "return_attn_probs",
|
| 165 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 166 |
+
"default": "False",
|
| 167 |
+
"annotation": null
|
| 168 |
+
}
|
| 169 |
+
],
|
| 170 |
+
"return_annotation": null
|
| 171 |
+
}
|
| 172 |
+
},
|
| 173 |
+
{
|
| 174 |
+
"name": "flash_attn_qkvpacked_func",
|
| 175 |
+
"qualname": "flash_attn_qkvpacked_func",
|
| 176 |
+
"module": "_flash_attn3_cuda_fe2bb66.flash_attn_interface",
|
| 177 |
+
"kind": "function",
|
| 178 |
+
"doc": "dropout_p should be set to 0.0 during evaluation\nIf Q, K, V are already stacked into 1 tensor, this function will be faster than\ncalling flash_attn_func on Q, K, V since the backward pass avoids explicit concatenation\nof the gradients of Q, K, V.\nFor multi-query and grouped-query attention (MQA/GQA), please see\nflash_attn_kvpacked_func and flash_attn_func.\n\nIf window_size != (-1, -1), implements sliding window local attention. Query at position i\nwill only attend to keys between [i - window_size[0], i + window_size[1]] inclusive.\n\nArguments:\n qkv: (batch_size, seqlen, 3, nheads, headdim)\n dropout_p: float. Dropout probability.\n softmax_scale: float. The scaling of QK^T before applying softmax.\n Default to 1 / sqrt(headdim).\n causal: bool. Whether to apply causal attention mask (e.g., for auto-regressive modeling).\n window_size: (left, right). If not (-1, -1), implements sliding window local attention.\n softcap: float. Anything > 0 activates softcapping attention.\n alibi_slopes: (nheads,) or (batch_size, nheads), fp32. A bias of (-alibi_slope * |i - j|) is added to\n the attention score of query i and key j.\n deterministic: bool. Whether to use the deterministic implementation of the backward pass,\n which is slightly slower and uses more memory. The forward pass is always deterministic.\n return_attn_probs: bool. Whether to return the attention probabilities. This option is for\n testing only. The returned probabilities are not guaranteed to be correct\n (they might not have the right scaling).\nReturn:\n out: (batch_size, seqlen, nheads, headdim).\n softmax_lse [optional, if return_attn_probs=True]: (batch_size, nheads, seqlen). The\n logsumexp of each row of the matrix QK^T * scaling (e.g., log of the softmax\n normalization factor).\n S_dmask [optional, if return_attn_probs=True]: (batch_size, nheads, seqlen, seqlen).\n The output of softmax (possibly with different scaling). It also encodes the dropout\n pattern (negative means that location was dropped, nonnegative means it was kept).",
|
| 179 |
+
"signature": {
|
| 180 |
+
"parameters": [
|
| 181 |
+
{
|
| 182 |
+
"name": "qkv",
|
| 183 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 184 |
+
"default": null,
|
| 185 |
+
"annotation": null
|
| 186 |
+
},
|
| 187 |
+
{
|
| 188 |
+
"name": "softmax_scale",
|
| 189 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 190 |
+
"default": "None",
|
| 191 |
+
"annotation": null
|
| 192 |
+
},
|
| 193 |
+
{
|
| 194 |
+
"name": "causal",
|
| 195 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 196 |
+
"default": "False",
|
| 197 |
+
"annotation": null
|
| 198 |
+
},
|
| 199 |
+
{
|
| 200 |
+
"name": "q_descale",
|
| 201 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 202 |
+
"default": "None",
|
| 203 |
+
"annotation": null
|
| 204 |
+
},
|
| 205 |
+
{
|
| 206 |
+
"name": "k_descale",
|
| 207 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 208 |
+
"default": "None",
|
| 209 |
+
"annotation": null
|
| 210 |
+
},
|
| 211 |
+
{
|
| 212 |
+
"name": "v_descale",
|
| 213 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 214 |
+
"default": "None",
|
| 215 |
+
"annotation": null
|
| 216 |
+
},
|
| 217 |
+
{
|
| 218 |
+
"name": "window_size",
|
| 219 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 220 |
+
"default": "(-1, -1)",
|
| 221 |
+
"annotation": null
|
| 222 |
+
},
|
| 223 |
+
{
|
| 224 |
+
"name": "attention_chunk",
|
| 225 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 226 |
+
"default": "0",
|
| 227 |
+
"annotation": null
|
| 228 |
+
},
|
| 229 |
+
{
|
| 230 |
+
"name": "softcap",
|
| 231 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 232 |
+
"default": "0.0",
|
| 233 |
+
"annotation": null
|
| 234 |
+
},
|
| 235 |
+
{
|
| 236 |
+
"name": "deterministic",
|
| 237 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 238 |
+
"default": "False",
|
| 239 |
+
"annotation": null
|
| 240 |
+
},
|
| 241 |
+
{
|
| 242 |
+
"name": "num_heads_q",
|
| 243 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 244 |
+
"default": "None",
|
| 245 |
+
"annotation": null
|
| 246 |
+
},
|
| 247 |
+
{
|
| 248 |
+
"name": "sm_margin",
|
| 249 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 250 |
+
"default": "0",
|
| 251 |
+
"annotation": null
|
| 252 |
+
},
|
| 253 |
+
{
|
| 254 |
+
"name": "return_attn_probs",
|
| 255 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 256 |
+
"default": "False",
|
| 257 |
+
"annotation": null
|
| 258 |
+
}
|
| 259 |
+
],
|
| 260 |
+
"return_annotation": null
|
| 261 |
+
}
|
| 262 |
+
},
|
| 263 |
+
{
|
| 264 |
+
"name": "flash_attn_varlen_func",
|
| 265 |
+
"qualname": "flash_attn_varlen_func",
|
| 266 |
+
"module": "_flash_attn3_cuda_fe2bb66.flash_attn_interface",
|
| 267 |
+
"kind": "function",
|
| 268 |
+
"doc": null,
|
| 269 |
+
"signature": {
|
| 270 |
+
"parameters": [
|
| 271 |
+
{
|
| 272 |
+
"name": "q",
|
| 273 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 274 |
+
"default": null,
|
| 275 |
+
"annotation": null
|
| 276 |
+
},
|
| 277 |
+
{
|
| 278 |
+
"name": "k",
|
| 279 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 280 |
+
"default": null,
|
| 281 |
+
"annotation": null
|
| 282 |
+
},
|
| 283 |
+
{
|
| 284 |
+
"name": "v",
|
| 285 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 286 |
+
"default": null,
|
| 287 |
+
"annotation": null
|
| 288 |
+
},
|
| 289 |
+
{
|
| 290 |
+
"name": "cu_seqlens_q",
|
| 291 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 292 |
+
"default": null,
|
| 293 |
+
"annotation": null
|
| 294 |
+
},
|
| 295 |
+
{
|
| 296 |
+
"name": "cu_seqlens_k",
|
| 297 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 298 |
+
"default": null,
|
| 299 |
+
"annotation": null
|
| 300 |
+
},
|
| 301 |
+
{
|
| 302 |
+
"name": "max_seqlen_q",
|
| 303 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 304 |
+
"default": null,
|
| 305 |
+
"annotation": null
|
| 306 |
+
},
|
| 307 |
+
{
|
| 308 |
+
"name": "max_seqlen_k",
|
| 309 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 310 |
+
"default": null,
|
| 311 |
+
"annotation": null
|
| 312 |
+
},
|
| 313 |
+
{
|
| 314 |
+
"name": "seqused_q",
|
| 315 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 316 |
+
"default": "None",
|
| 317 |
+
"annotation": null
|
| 318 |
+
},
|
| 319 |
+
{
|
| 320 |
+
"name": "seqused_k",
|
| 321 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 322 |
+
"default": "None",
|
| 323 |
+
"annotation": null
|
| 324 |
+
},
|
| 325 |
+
{
|
| 326 |
+
"name": "softmax_scale",
|
| 327 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 328 |
+
"default": "None",
|
| 329 |
+
"annotation": null
|
| 330 |
+
},
|
| 331 |
+
{
|
| 332 |
+
"name": "causal",
|
| 333 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 334 |
+
"default": "False",
|
| 335 |
+
"annotation": null
|
| 336 |
+
},
|
| 337 |
+
{
|
| 338 |
+
"name": "qv",
|
| 339 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 340 |
+
"default": "None",
|
| 341 |
+
"annotation": null
|
| 342 |
+
},
|
| 343 |
+
{
|
| 344 |
+
"name": "q_descale",
|
| 345 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 346 |
+
"default": "None",
|
| 347 |
+
"annotation": null
|
| 348 |
+
},
|
| 349 |
+
{
|
| 350 |
+
"name": "k_descale",
|
| 351 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 352 |
+
"default": "None",
|
| 353 |
+
"annotation": null
|
| 354 |
+
},
|
| 355 |
+
{
|
| 356 |
+
"name": "v_descale",
|
| 357 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 358 |
+
"default": "None",
|
| 359 |
+
"annotation": null
|
| 360 |
+
},
|
| 361 |
+
{
|
| 362 |
+
"name": "window_size",
|
| 363 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 364 |
+
"default": "(-1, -1)",
|
| 365 |
+
"annotation": null
|
| 366 |
+
},
|
| 367 |
+
{
|
| 368 |
+
"name": "attention_chunk",
|
| 369 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 370 |
+
"default": "0",
|
| 371 |
+
"annotation": null
|
| 372 |
+
},
|
| 373 |
+
{
|
| 374 |
+
"name": "softcap",
|
| 375 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 376 |
+
"default": "0.0",
|
| 377 |
+
"annotation": null
|
| 378 |
+
},
|
| 379 |
+
{
|
| 380 |
+
"name": "num_splits",
|
| 381 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 382 |
+
"default": "1",
|
| 383 |
+
"annotation": null
|
| 384 |
+
},
|
| 385 |
+
{
|
| 386 |
+
"name": "pack_gqa",
|
| 387 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 388 |
+
"default": "None",
|
| 389 |
+
"annotation": null
|
| 390 |
+
},
|
| 391 |
+
{
|
| 392 |
+
"name": "deterministic",
|
| 393 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 394 |
+
"default": "False",
|
| 395 |
+
"annotation": null
|
| 396 |
+
},
|
| 397 |
+
{
|
| 398 |
+
"name": "sm_margin",
|
| 399 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 400 |
+
"default": "0",
|
| 401 |
+
"annotation": null
|
| 402 |
+
},
|
| 403 |
+
{
|
| 404 |
+
"name": "return_attn_probs",
|
| 405 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 406 |
+
"default": "False",
|
| 407 |
+
"annotation": null
|
| 408 |
+
}
|
| 409 |
+
],
|
| 410 |
+
"return_annotation": null
|
| 411 |
+
}
|
| 412 |
+
},
|
| 413 |
+
{
|
| 414 |
+
"name": "flash_attn_with_kvcache",
|
| 415 |
+
"qualname": "flash_attn_with_kvcache",
|
| 416 |
+
"module": "_flash_attn3_cuda_fe2bb66.flash_attn_interface",
|
| 417 |
+
"kind": "function",
|
| 418 |
+
"doc": "If k and v are not None, k_cache and v_cache will be updated *inplace* with the new values from\nk and v. This is useful for incremental decoding: you can pass in the cached keys/values from\nthe previous step, and update them with the new keys/values from the current step, and do\nattention with the updated cache, all in 1 kernel.\n\nIf you pass in k / v, you must make sure that the cache is large enough to hold the new values.\nFor example, the KV cache could be pre-allocated with the max sequence length, and you can use\ncache_seqlens to keep track of the current sequence lengths of each sequence in the batch.\n\nAlso apply rotary embedding if rotary_cos and rotary_sin are passed in. The key @k will be\nrotated by rotary_cos and rotary_sin at indices cache_seqlens, cache_seqlens + 1, etc.\nIf causal or local (i.e., window_size != (-1, -1)), the query @q will be rotated by rotary_cos\nand rotary_sin at indices cache_seqlens, cache_seqlens + 1, etc.\nIf not causal and not local, the query @q will be rotated by rotary_cos and rotary_sin at\nindices cache_seqlens only (i.e. we consider all tokens in @q to be at position cache_seqlens).\n\nSee tests/test_flash_attn.py::test_flash_attn_kvcache for examples of how to use this function.\n\nSupports multi-query and grouped-query attention (MQA/GQA) by passing in KV with fewer heads\nthan Q. Note that the number of heads in Q must be divisible by the number of heads in KV.\nFor example, if Q has 6 heads and K, V have 2 heads, head 0, 1, 2 of Q will attention to head\n0 of K, V, and head 3, 4, 5 of Q will attention to head 1 of K, V.\n\nIf causal=True, the causal mask is aligned to the bottom right corner of the attention matrix.\nFor example, if seqlen_q = 2 and seqlen_k = 5, the causal mask (1 = keep, 0 = masked out) is:\n 1 1 1 1 0\n 1 1 1 1 1\nIf seqlen_q = 5 and seqlen_k = 2, the causal mask is:\n 0 0\n 0 0\n 0 0\n 1 0\n 1 1\nIf the row of the mask is all zero, the output will be zero.\n\nIf window_size != (-1, -1), implements sliding window local attention. Query at position i\nwill only attend to keys between\n[i + seqlen_k - seqlen_q - window_size[0], i + seqlen_k - seqlen_q + window_size[1]] inclusive.\n\nNote: Does not support backward pass.\n\nArguments:\n q: (batch_size, seqlen, nheads, headdim)\n k_cache: (batch_size_cache, seqlen_cache, nheads_k, headdim) if there's no page_table,\n or (num_blocks, page_block_size, nheads_k, headdim) if there's a page_table (i.e. paged KV cache)\n page_block_size can be arbitrary (e.g, 1, 2, 3, 64, etc.).\n v_cache: (batch_size_cache, seqlen_cache, nheads_k, headdim_v) if there's no page_table,\n or (num_blocks, page_block_size, nheads_k, headdim_v) if there's a page_table (i.e. paged KV cache)\n k [optional]: (batch_size, seqlen_new, nheads_k, headdim). If not None, we concatenate\n k with k_cache, starting at the indices specified by cache_seqlens.\n v [optional]: (batch_size, seqlen_new, nheads_k, headdim_v). Similar to k.\n qv [optional]: (batch_size, seqlen, nheads, headdim_v)\n rotary_cos [optional]: (seqlen_ro, rotary_dim / 2). If not None, we apply rotary embedding\n to k and q. Only applicable if k and v are passed in. rotary_dim must be divisible by 16.\n rotary_sin [optional]: (seqlen_ro, rotary_dim / 2). Similar to rotary_cos.\n cache_seqlens: int, or (batch_size,), dtype torch.int32. The sequence lengths of the\n KV cache.\n cache_batch_idx: (batch_size,), dtype torch.int32. The indices used to index into the KV cache.\n If None, we assume that the batch indices are [0, 1, 2, ..., batch_size - 1].\n If the indices are not distinct, and k and v are provided, the values updated in the cache\n might come from any of the duplicate indices.\n cache_leftpad: (batch_size,), dtype torch.int32. The index that the KV cache starts. If None, assume 0.\n page_table [optional]: (batch_size, max_num_blocks_per_seq), dtype torch.int32.\n softmax_scale: float. The scaling of QK^T before applying softmax.\n Default to 1 / sqrt(headdim).\n causal: bool. Whether to apply causal attention mask (e.g., for auto-regressive modeling).\n window_size: (left, right). If not (-1, -1), implements sliding window local attention.\n softcap: float. Anything > 0 activates softcapping attention.\n rotary_interleaved: bool. Only applicable if rotary_cos and rotary_sin are passed in.\n If True, rotary embedding will combine dimensions 0 & 1, 2 & 3, etc. If False,\n rotary embedding will combine dimensions 0 & rotary_dim / 2, 1 & rotary_dim / 2 + 1\n (i.e. GPT-NeoX style).\n num_splits: int. If > 1, split the key/value into this many chunks along the sequence.\n If num_splits == 1, we don't split the key/value. If num_splits == 0, we use a heuristic\n to automatically determine the number of splits.\n Don't change this unless you know what you are doing.\n return_softmax_lse: bool. Whether to return the logsumexp of the attention scores.\n\nReturn:\n out: (batch_size, seqlen, nheads, headdim).\n softmax_lse [optional, if return_softmax_lse=True]: (batch_size, nheads, seqlen). The\n logsumexp of each row of the matrix QK^T * scaling (e.g., log of the softmax\n normalization factor).",
|
| 419 |
+
"signature": {
|
| 420 |
+
"parameters": [
|
| 421 |
+
{
|
| 422 |
+
"name": "q",
|
| 423 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 424 |
+
"default": null,
|
| 425 |
+
"annotation": null
|
| 426 |
+
},
|
| 427 |
+
{
|
| 428 |
+
"name": "k_cache",
|
| 429 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 430 |
+
"default": null,
|
| 431 |
+
"annotation": null
|
| 432 |
+
},
|
| 433 |
+
{
|
| 434 |
+
"name": "v_cache",
|
| 435 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 436 |
+
"default": null,
|
| 437 |
+
"annotation": null
|
| 438 |
+
},
|
| 439 |
+
{
|
| 440 |
+
"name": "k",
|
| 441 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 442 |
+
"default": "None",
|
| 443 |
+
"annotation": null
|
| 444 |
+
},
|
| 445 |
+
{
|
| 446 |
+
"name": "v",
|
| 447 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 448 |
+
"default": "None",
|
| 449 |
+
"annotation": null
|
| 450 |
+
},
|
| 451 |
+
{
|
| 452 |
+
"name": "qv",
|
| 453 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 454 |
+
"default": "None",
|
| 455 |
+
"annotation": null
|
| 456 |
+
},
|
| 457 |
+
{
|
| 458 |
+
"name": "rotary_cos",
|
| 459 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 460 |
+
"default": "None",
|
| 461 |
+
"annotation": null
|
| 462 |
+
},
|
| 463 |
+
{
|
| 464 |
+
"name": "rotary_sin",
|
| 465 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 466 |
+
"default": "None",
|
| 467 |
+
"annotation": null
|
| 468 |
+
},
|
| 469 |
+
{
|
| 470 |
+
"name": "cache_seqlens",
|
| 471 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 472 |
+
"default": "None",
|
| 473 |
+
"annotation": "int | torch.Tensor | None"
|
| 474 |
+
},
|
| 475 |
+
{
|
| 476 |
+
"name": "cache_batch_idx",
|
| 477 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 478 |
+
"default": "None",
|
| 479 |
+
"annotation": "torch.Tensor | None"
|
| 480 |
+
},
|
| 481 |
+
{
|
| 482 |
+
"name": "cache_leftpad",
|
| 483 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 484 |
+
"default": "None",
|
| 485 |
+
"annotation": "torch.Tensor | None"
|
| 486 |
+
},
|
| 487 |
+
{
|
| 488 |
+
"name": "page_table",
|
| 489 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 490 |
+
"default": "None",
|
| 491 |
+
"annotation": "torch.Tensor | None"
|
| 492 |
+
},
|
| 493 |
+
{
|
| 494 |
+
"name": "cu_seqlens_q",
|
| 495 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 496 |
+
"default": "None",
|
| 497 |
+
"annotation": "torch.Tensor | None"
|
| 498 |
+
},
|
| 499 |
+
{
|
| 500 |
+
"name": "cu_seqlens_k_new",
|
| 501 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 502 |
+
"default": "None",
|
| 503 |
+
"annotation": "torch.Tensor | None"
|
| 504 |
+
},
|
| 505 |
+
{
|
| 506 |
+
"name": "max_seqlen_q",
|
| 507 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 508 |
+
"default": "None",
|
| 509 |
+
"annotation": "int | None"
|
| 510 |
+
},
|
| 511 |
+
{
|
| 512 |
+
"name": "rotary_seqlens",
|
| 513 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 514 |
+
"default": "None",
|
| 515 |
+
"annotation": "torch.Tensor | None"
|
| 516 |
+
},
|
| 517 |
+
{
|
| 518 |
+
"name": "q_descale",
|
| 519 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 520 |
+
"default": "None",
|
| 521 |
+
"annotation": "torch.Tensor | None"
|
| 522 |
+
},
|
| 523 |
+
{
|
| 524 |
+
"name": "k_descale",
|
| 525 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 526 |
+
"default": "None",
|
| 527 |
+
"annotation": "torch.Tensor | None"
|
| 528 |
+
},
|
| 529 |
+
{
|
| 530 |
+
"name": "v_descale",
|
| 531 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 532 |
+
"default": "None",
|
| 533 |
+
"annotation": "torch.Tensor | None"
|
| 534 |
+
},
|
| 535 |
+
{
|
| 536 |
+
"name": "softmax_scale",
|
| 537 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 538 |
+
"default": "None",
|
| 539 |
+
"annotation": null
|
| 540 |
+
},
|
| 541 |
+
{
|
| 542 |
+
"name": "causal",
|
| 543 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 544 |
+
"default": "False",
|
| 545 |
+
"annotation": null
|
| 546 |
+
},
|
| 547 |
+
{
|
| 548 |
+
"name": "window_size",
|
| 549 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 550 |
+
"default": "(-1, -1)",
|
| 551 |
+
"annotation": null
|
| 552 |
+
},
|
| 553 |
+
{
|
| 554 |
+
"name": "attention_chunk",
|
| 555 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 556 |
+
"default": "0",
|
| 557 |
+
"annotation": null
|
| 558 |
+
},
|
| 559 |
+
{
|
| 560 |
+
"name": "softcap",
|
| 561 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 562 |
+
"default": "0.0",
|
| 563 |
+
"annotation": null
|
| 564 |
+
},
|
| 565 |
+
{
|
| 566 |
+
"name": "rotary_interleaved",
|
| 567 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 568 |
+
"default": "True",
|
| 569 |
+
"annotation": null
|
| 570 |
+
},
|
| 571 |
+
{
|
| 572 |
+
"name": "scheduler_metadata",
|
| 573 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 574 |
+
"default": "None",
|
| 575 |
+
"annotation": null
|
| 576 |
+
},
|
| 577 |
+
{
|
| 578 |
+
"name": "num_splits",
|
| 579 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 580 |
+
"default": "0",
|
| 581 |
+
"annotation": null
|
| 582 |
+
},
|
| 583 |
+
{
|
| 584 |
+
"name": "pack_gqa",
|
| 585 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 586 |
+
"default": "None",
|
| 587 |
+
"annotation": null
|
| 588 |
+
},
|
| 589 |
+
{
|
| 590 |
+
"name": "sm_margin",
|
| 591 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 592 |
+
"default": "0",
|
| 593 |
+
"annotation": null
|
| 594 |
+
},
|
| 595 |
+
{
|
| 596 |
+
"name": "return_softmax_lse",
|
| 597 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 598 |
+
"default": "False",
|
| 599 |
+
"annotation": null
|
| 600 |
+
}
|
| 601 |
+
],
|
| 602 |
+
"return_annotation": null
|
| 603 |
+
}
|
| 604 |
+
},
|
| 605 |
+
{
|
| 606 |
+
"name": "get_scheduler_metadata",
|
| 607 |
+
"qualname": "get_scheduler_metadata",
|
| 608 |
+
"module": "_flash_attn3_cuda_fe2bb66.flash_attn_interface",
|
| 609 |
+
"kind": "function",
|
| 610 |
+
"doc": null,
|
| 611 |
+
"signature": {
|
| 612 |
+
"parameters": [
|
| 613 |
+
{
|
| 614 |
+
"name": "batch_size",
|
| 615 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 616 |
+
"default": null,
|
| 617 |
+
"annotation": null
|
| 618 |
+
},
|
| 619 |
+
{
|
| 620 |
+
"name": "max_seqlen_q",
|
| 621 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 622 |
+
"default": null,
|
| 623 |
+
"annotation": null
|
| 624 |
+
},
|
| 625 |
+
{
|
| 626 |
+
"name": "max_seqlen_k",
|
| 627 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 628 |
+
"default": null,
|
| 629 |
+
"annotation": null
|
| 630 |
+
},
|
| 631 |
+
{
|
| 632 |
+
"name": "num_heads_q",
|
| 633 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 634 |
+
"default": null,
|
| 635 |
+
"annotation": null
|
| 636 |
+
},
|
| 637 |
+
{
|
| 638 |
+
"name": "num_heads_kv",
|
| 639 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 640 |
+
"default": null,
|
| 641 |
+
"annotation": null
|
| 642 |
+
},
|
| 643 |
+
{
|
| 644 |
+
"name": "headdim",
|
| 645 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 646 |
+
"default": null,
|
| 647 |
+
"annotation": null
|
| 648 |
+
},
|
| 649 |
+
{
|
| 650 |
+
"name": "cache_seqlens",
|
| 651 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 652 |
+
"default": null,
|
| 653 |
+
"annotation": "torch.Tensor"
|
| 654 |
+
},
|
| 655 |
+
{
|
| 656 |
+
"name": "qkv_dtype",
|
| 657 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 658 |
+
"default": "torch.bfloat16",
|
| 659 |
+
"annotation": null
|
| 660 |
+
},
|
| 661 |
+
{
|
| 662 |
+
"name": "headdim_v",
|
| 663 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 664 |
+
"default": "None",
|
| 665 |
+
"annotation": null
|
| 666 |
+
},
|
| 667 |
+
{
|
| 668 |
+
"name": "cu_seqlens_q",
|
| 669 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 670 |
+
"default": "None",
|
| 671 |
+
"annotation": "torch.Tensor | None"
|
| 672 |
+
},
|
| 673 |
+
{
|
| 674 |
+
"name": "cu_seqlens_k_new",
|
| 675 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 676 |
+
"default": "None",
|
| 677 |
+
"annotation": "torch.Tensor | None"
|
| 678 |
+
},
|
| 679 |
+
{
|
| 680 |
+
"name": "cache_leftpad",
|
| 681 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 682 |
+
"default": "None",
|
| 683 |
+
"annotation": "torch.Tensor | None"
|
| 684 |
+
},
|
| 685 |
+
{
|
| 686 |
+
"name": "page_size",
|
| 687 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 688 |
+
"default": "None",
|
| 689 |
+
"annotation": "int | None"
|
| 690 |
+
},
|
| 691 |
+
{
|
| 692 |
+
"name": "max_seqlen_k_new",
|
| 693 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 694 |
+
"default": "0",
|
| 695 |
+
"annotation": null
|
| 696 |
+
},
|
| 697 |
+
{
|
| 698 |
+
"name": "causal",
|
| 699 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 700 |
+
"default": "False",
|
| 701 |
+
"annotation": null
|
| 702 |
+
},
|
| 703 |
+
{
|
| 704 |
+
"name": "window_size",
|
| 705 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 706 |
+
"default": "(-1, -1)",
|
| 707 |
+
"annotation": null
|
| 708 |
+
},
|
| 709 |
+
{
|
| 710 |
+
"name": "attention_chunk",
|
| 711 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 712 |
+
"default": "0",
|
| 713 |
+
"annotation": null
|
| 714 |
+
},
|
| 715 |
+
{
|
| 716 |
+
"name": "has_softcap",
|
| 717 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 718 |
+
"default": "False",
|
| 719 |
+
"annotation": null
|
| 720 |
+
},
|
| 721 |
+
{
|
| 722 |
+
"name": "num_splits",
|
| 723 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 724 |
+
"default": "0",
|
| 725 |
+
"annotation": null
|
| 726 |
+
},
|
| 727 |
+
{
|
| 728 |
+
"name": "pack_gqa",
|
| 729 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 730 |
+
"default": "None",
|
| 731 |
+
"annotation": null
|
| 732 |
+
},
|
| 733 |
+
{
|
| 734 |
+
"name": "sm_margin",
|
| 735 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 736 |
+
"default": "0",
|
| 737 |
+
"annotation": null
|
| 738 |
+
}
|
| 739 |
+
],
|
| 740 |
+
"return_annotation": null
|
| 741 |
+
}
|
| 742 |
+
}
|
| 743 |
+
],
|
| 744 |
+
"layers": []
|
| 745 |
+
}
|
build/torch-stable-abi29-cu128-x86_64-linux/__init__.py
ADDED
|
@@ -0,0 +1,20 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
from ._ops import add_op_namespace_prefix, ops
|
| 2 |
+
from .flash_attn_interface import (
|
| 3 |
+
flash_attn_combine,
|
| 4 |
+
flash_attn_func,
|
| 5 |
+
flash_attn_qkvpacked_func,
|
| 6 |
+
flash_attn_varlen_func,
|
| 7 |
+
flash_attn_with_kvcache,
|
| 8 |
+
get_scheduler_metadata,
|
| 9 |
+
)
|
| 10 |
+
|
| 11 |
+
__all__ = [
|
| 12 |
+
"add_op_namespace_prefix",
|
| 13 |
+
"flash_attn_combine",
|
| 14 |
+
"flash_attn_func",
|
| 15 |
+
"flash_attn_qkvpacked_func",
|
| 16 |
+
"flash_attn_varlen_func",
|
| 17 |
+
"flash_attn_with_kvcache",
|
| 18 |
+
"get_scheduler_metadata",
|
| 19 |
+
"ops",
|
| 20 |
+
]
|
build/{torch-stable-abi29-cu130-x86_64-linux/_flash_attn3_cuda_90fb2c3.abi3.so → torch-stable-abi29-cu128-x86_64-linux/_flash_attn3_cuda_fe2bb66.abi3.so}
RENAMED
|
@@ -1,3 +1,3 @@
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
-
oid sha256:
|
| 3 |
-
size
|
|
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:1bb527c616970eb30b22e9424bf71e4a889a014f318159531f9af4f3daa77c94
|
| 3 |
+
size 530776216
|
build/torch-stable-abi29-cu128-x86_64-linux/_ops.py
ADDED
|
@@ -0,0 +1,9 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import torch
|
| 2 |
+
from . import _flash_attn3_cuda_fe2bb66
|
| 3 |
+
ops = torch.ops._flash_attn3_cuda_fe2bb66
|
| 4 |
+
|
| 5 |
+
def add_op_namespace_prefix(op_name: str):
|
| 6 |
+
"""
|
| 7 |
+
Prefix op by namespace.
|
| 8 |
+
"""
|
| 9 |
+
return f"_flash_attn3_cuda_fe2bb66::{op_name}"
|
build/torch-stable-abi29-cu128-x86_64-linux/flash_attn_config.py
ADDED
|
@@ -0,0 +1,7 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Auto-generated by flash attention 3 setup.py
|
| 2 |
+
CONFIG = {'build_flags': {'FLASHATTENTION_DISABLE_BACKWARD': False, 'FLASHATTENTION_DISABLE_SPLIT': False, 'FLASHATTENTION_DISABLE_PAGEDKV': False, 'FLASHATTENTION_DISABLE_APPENDKV': False, 'FLASHATTENTION_DISABLE_LOCAL': False, 'FLASHATTENTION_DISABLE_SOFTCAP': False, 'FLASHATTENTION_DISABLE_PACKGQA': False, 'FLASHATTENTION_DISABLE_FP16': False, 'FLASHATTENTION_DISABLE_FP8': False, 'FLASHATTENTION_DISABLE_VARLEN': False, 'FLASHATTENTION_DISABLE_CLUSTER': False, 'FLASHATTENTION_DISABLE_HDIM64': False, 'FLASHATTENTION_DISABLE_HDIM96': False, 'FLASHATTENTION_DISABLE_HDIM128': False, 'FLASHATTENTION_DISABLE_HDIM192': False, 'FLASHATTENTION_DISABLE_HDIM256': False, 'FLASHATTENTION_DISABLE_SM8x': False, 'FLASHATTENTION_ENABLE_VCOLMAJOR': False, 'FLASH_ATTENTION_DISABLE_HDIMDIFF64': False, 'FLASH_ATTENTION_DISABLE_HDIMDIFF192': False}}
|
| 3 |
+
|
| 4 |
+
def show():
|
| 5 |
+
from pprint import pprint
|
| 6 |
+
pprint(CONFIG)
|
| 7 |
+
|
build/torch-stable-abi29-cu128-x86_64-linux/flash_attn_interface.py
ADDED
|
@@ -0,0 +1,1127 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Copyright (c) 2023, Tri Dao.
|
| 2 |
+
|
| 3 |
+
from typing import Optional, Union, List, Tuple
|
| 4 |
+
|
| 5 |
+
import torch
|
| 6 |
+
import torch.nn as nn
|
| 7 |
+
|
| 8 |
+
from ._ops import ops as flash_attn_3_cuda
|
| 9 |
+
from ._ops import add_op_namespace_prefix
|
| 10 |
+
|
| 11 |
+
def maybe_contiguous(x):
|
| 12 |
+
return x.contiguous() if x is not None and x.stride(-1) != 1 else x
|
| 13 |
+
|
| 14 |
+
|
| 15 |
+
def round_multiple(x, m):
|
| 16 |
+
return (x + m - 1) // m * m
|
| 17 |
+
|
| 18 |
+
|
| 19 |
+
def round_up_headdim(head_size: int) -> int:
|
| 20 |
+
from .flash_attn_config import CONFIG
|
| 21 |
+
|
| 22 |
+
if not CONFIG["build_flags"]["FLASHATTENTION_DISABLE_HDIM64"]:
|
| 23 |
+
if head_size <= 64:
|
| 24 |
+
return 64
|
| 25 |
+
if not CONFIG["build_flags"]["FLASHATTENTION_DISABLE_HDIM96"]:
|
| 26 |
+
if head_size <= 96:
|
| 27 |
+
return 96
|
| 28 |
+
if not CONFIG["build_flags"]["FLASHATTENTION_DISABLE_HDIM128"]:
|
| 29 |
+
if head_size <= 128:
|
| 30 |
+
return 128
|
| 31 |
+
if not CONFIG["build_flags"]["FLASHATTENTION_DISABLE_HDIM192"]:
|
| 32 |
+
if head_size <= 192:
|
| 33 |
+
return 192
|
| 34 |
+
if not CONFIG["build_flags"]["FLASHATTENTION_DISABLE_HDIM256"]:
|
| 35 |
+
if head_size <= 256:
|
| 36 |
+
return 256
|
| 37 |
+
return 256
|
| 38 |
+
|
| 39 |
+
|
| 40 |
+
@torch.library.custom_op(add_op_namespace_prefix("_flash_attn_forward"), mutates_args=(), device_types="cuda")
|
| 41 |
+
def _flash_attn_forward(
|
| 42 |
+
q: torch.Tensor,
|
| 43 |
+
k: torch.Tensor,
|
| 44 |
+
v: torch.Tensor,
|
| 45 |
+
k_new: Optional[torch.Tensor] = None,
|
| 46 |
+
v_new: Optional[torch.Tensor] = None,
|
| 47 |
+
qv: Optional[torch.Tensor] = None,
|
| 48 |
+
out_: Optional[torch.Tensor] = None,
|
| 49 |
+
cu_seqlens_q: Optional[torch.Tensor] = None,
|
| 50 |
+
cu_seqlens_k: Optional[torch.Tensor] = None,
|
| 51 |
+
cu_seqlens_k_new: Optional[torch.Tensor] = None,
|
| 52 |
+
seqused_q: Optional[torch.Tensor] = None,
|
| 53 |
+
seqused_k: Optional[torch.Tensor] = None,
|
| 54 |
+
max_seqlen_q: Optional[int] = None,
|
| 55 |
+
max_seqlen_k: Optional[int] = None,
|
| 56 |
+
page_table: Optional[torch.Tensor] = None,
|
| 57 |
+
kv_batch_idx: Optional[torch.Tensor] = None,
|
| 58 |
+
leftpad_k: Optional[torch.Tensor] = None,
|
| 59 |
+
rotary_cos: Optional[torch.Tensor] = None,
|
| 60 |
+
rotary_sin: Optional[torch.Tensor] = None,
|
| 61 |
+
seqlens_rotary: Optional[torch.Tensor] = None,
|
| 62 |
+
q_descale: Optional[torch.Tensor] = None,
|
| 63 |
+
k_descale: Optional[torch.Tensor] = None,
|
| 64 |
+
v_descale: Optional[torch.Tensor] = None,
|
| 65 |
+
softmax_scale: Optional[float] = None,
|
| 66 |
+
causal: bool = False,
|
| 67 |
+
window_size_left: int = -1,
|
| 68 |
+
window_size_right: int = -1,
|
| 69 |
+
attention_chunk: int = 0,
|
| 70 |
+
softcap: float = 0.0,
|
| 71 |
+
rotary_interleaved: bool = True,
|
| 72 |
+
scheduler_metadata: Optional[torch.Tensor] = None,
|
| 73 |
+
num_splits: int = 1,
|
| 74 |
+
pack_gqa: Optional[bool] = None,
|
| 75 |
+
sm_margin: int = 0,
|
| 76 |
+
) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
|
| 77 |
+
q, k, k_new, v_new = [maybe_contiguous(x) for x in (q, k, k_new, v_new)]
|
| 78 |
+
v = v.contiguous() if v.stride(-1) != 1 and v.stride(-3) != 1 else v
|
| 79 |
+
cu_seqlens_q, cu_seqlens_k, cu_seqlens_k_new = [
|
| 80 |
+
maybe_contiguous(x) for x in (cu_seqlens_q, cu_seqlens_k, cu_seqlens_k_new)
|
| 81 |
+
]
|
| 82 |
+
seqused_q, seqused_k = [maybe_contiguous(x) for x in (seqused_q, seqused_k)]
|
| 83 |
+
page_table, kv_batch_idx, leftpad_k = [
|
| 84 |
+
maybe_contiguous(x) for x in (page_table, kv_batch_idx, leftpad_k)
|
| 85 |
+
]
|
| 86 |
+
rotary_cos, rotary_sin = [maybe_contiguous(x) for x in (rotary_cos, rotary_sin)]
|
| 87 |
+
seqlens_rotary = maybe_contiguous(seqlens_rotary)
|
| 88 |
+
out, softmax_lse, out_accum, softmax_lse_accum = flash_attn_3_cuda.fwd(
|
| 89 |
+
q,
|
| 90 |
+
k,
|
| 91 |
+
v,
|
| 92 |
+
k_new,
|
| 93 |
+
v_new,
|
| 94 |
+
qv,
|
| 95 |
+
out_,
|
| 96 |
+
cu_seqlens_q,
|
| 97 |
+
cu_seqlens_k,
|
| 98 |
+
cu_seqlens_k_new,
|
| 99 |
+
seqused_q,
|
| 100 |
+
seqused_k,
|
| 101 |
+
max_seqlen_q,
|
| 102 |
+
max_seqlen_k,
|
| 103 |
+
page_table,
|
| 104 |
+
kv_batch_idx,
|
| 105 |
+
leftpad_k,
|
| 106 |
+
rotary_cos,
|
| 107 |
+
rotary_sin,
|
| 108 |
+
seqlens_rotary,
|
| 109 |
+
q_descale,
|
| 110 |
+
k_descale,
|
| 111 |
+
v_descale,
|
| 112 |
+
softmax_scale,
|
| 113 |
+
causal,
|
| 114 |
+
window_size_left,
|
| 115 |
+
window_size_right,
|
| 116 |
+
attention_chunk,
|
| 117 |
+
softcap,
|
| 118 |
+
rotary_interleaved,
|
| 119 |
+
scheduler_metadata,
|
| 120 |
+
num_splits,
|
| 121 |
+
pack_gqa,
|
| 122 |
+
sm_margin,
|
| 123 |
+
)
|
| 124 |
+
|
| 125 |
+
if out_accum is None:
|
| 126 |
+
out_accum = torch.tensor([], device=out.device)
|
| 127 |
+
|
| 128 |
+
if softmax_lse_accum is None:
|
| 129 |
+
softmax_lse_accum = torch.tensor([], device=out.device)
|
| 130 |
+
|
| 131 |
+
return out, softmax_lse, out_accum, softmax_lse_accum
|
| 132 |
+
|
| 133 |
+
|
| 134 |
+
@torch.library.register_fake(add_op_namespace_prefix("_flash_attn_forward"))
|
| 135 |
+
def _flash_attn_forward_fake(
|
| 136 |
+
q: torch.Tensor,
|
| 137 |
+
k: torch.Tensor,
|
| 138 |
+
v: torch.Tensor,
|
| 139 |
+
k_new: Optional[torch.Tensor] = None,
|
| 140 |
+
v_new: Optional[torch.Tensor] = None,
|
| 141 |
+
qv: Optional[torch.Tensor] = None,
|
| 142 |
+
out_: Optional[torch.Tensor] = None,
|
| 143 |
+
cu_seqlens_q: Optional[torch.Tensor] = None,
|
| 144 |
+
cu_seqlens_k: Optional[torch.Tensor] = None,
|
| 145 |
+
cu_seqlens_k_new: Optional[torch.Tensor] = None,
|
| 146 |
+
seqused_q: Optional[torch.Tensor] = None,
|
| 147 |
+
seqused_k: Optional[torch.Tensor] = None,
|
| 148 |
+
max_seqlen_q: Optional[int] = None,
|
| 149 |
+
max_seqlen_k: Optional[int] = None,
|
| 150 |
+
page_table: Optional[torch.Tensor] = None,
|
| 151 |
+
kv_batch_idx: Optional[torch.Tensor] = None,
|
| 152 |
+
leftpad_k: Optional[torch.Tensor] = None,
|
| 153 |
+
rotary_cos: Optional[torch.Tensor] = None,
|
| 154 |
+
rotary_sin: Optional[torch.Tensor] = None,
|
| 155 |
+
seqlens_rotary: Optional[torch.Tensor] = None,
|
| 156 |
+
q_descale: Optional[torch.Tensor] = None,
|
| 157 |
+
k_descale: Optional[torch.Tensor] = None,
|
| 158 |
+
v_descale: Optional[torch.Tensor] = None,
|
| 159 |
+
softmax_scale: Optional[float] = None,
|
| 160 |
+
causal: bool = False,
|
| 161 |
+
window_size_left: int = -1,
|
| 162 |
+
window_size_right: int = -1,
|
| 163 |
+
attention_chunk: int = 0,
|
| 164 |
+
softcap: float = 0.0,
|
| 165 |
+
rotary_interleaved: bool = True,
|
| 166 |
+
scheduler_metadata: Optional[torch.Tensor] = None,
|
| 167 |
+
num_splits: int = 1,
|
| 168 |
+
pack_gqa: Optional[bool] = None,
|
| 169 |
+
sm_margin: int = 0,
|
| 170 |
+
) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
|
| 171 |
+
"""
|
| 172 |
+
Symbolic fake implementation of flash attention forward.
|
| 173 |
+
Returns tensors with the correct shapes and dtypes without actual computation.
|
| 174 |
+
"""
|
| 175 |
+
|
| 176 |
+
# Determine if we're in varlen mode
|
| 177 |
+
is_varlen_q = cu_seqlens_q is not None
|
| 178 |
+
|
| 179 |
+
# Get dimensions from query tensor
|
| 180 |
+
if is_varlen_q:
|
| 181 |
+
# varlen mode: q is (total_q, num_heads, head_size)
|
| 182 |
+
total_q, num_heads, head_size = q.shape
|
| 183 |
+
batch_size = cu_seqlens_q.shape[0] - 1
|
| 184 |
+
|
| 185 |
+
if max_seqlen_q is None:
|
| 186 |
+
raise ValueError("max_seqlen_q must be provided if cu_seqlens_q is provided")
|
| 187 |
+
seqlen_q = max_seqlen_q
|
| 188 |
+
else:
|
| 189 |
+
# batch mode: q is (batch_size, seqlen_q, num_heads, head_size)
|
| 190 |
+
batch_size, seqlen_q, num_heads, head_size = q.shape
|
| 191 |
+
total_q = batch_size * q.shape[1]
|
| 192 |
+
# Get value head dimension
|
| 193 |
+
head_size_v = v.shape[-1]
|
| 194 |
+
|
| 195 |
+
# Determine output dtype (FP8 inputs produce BF16 outputs)
|
| 196 |
+
q_type = q.dtype
|
| 197 |
+
if q_type == torch.float8_e4m3fn:
|
| 198 |
+
out_dtype = torch.bfloat16
|
| 199 |
+
else:
|
| 200 |
+
out_dtype = q_type
|
| 201 |
+
|
| 202 |
+
# Create output tensor
|
| 203 |
+
if out_ is not None:
|
| 204 |
+
# If out_ is provided, _flash_attn_forward becomes non-functional
|
| 205 |
+
raise TypeError("Tracing (torch.compile/torch.export) with pre-allocated output tensor is not supported.")
|
| 206 |
+
|
| 207 |
+
if is_varlen_q:
|
| 208 |
+
out = torch.empty((total_q, num_heads, head_size_v), dtype=out_dtype, device=q.device)
|
| 209 |
+
else:
|
| 210 |
+
out = torch.empty((batch_size, seqlen_q, num_heads, head_size_v), dtype=out_dtype, device=q.device)
|
| 211 |
+
|
| 212 |
+
# Create softmax_lse tensor
|
| 213 |
+
if is_varlen_q:
|
| 214 |
+
softmax_lse = torch.empty((num_heads, total_q), dtype=torch.float32, device=q.device)
|
| 215 |
+
else:
|
| 216 |
+
softmax_lse = torch.empty((batch_size, num_heads, seqlen_q), dtype=torch.float32, device=q.device)
|
| 217 |
+
|
| 218 |
+
# TODO(guilhermeleobas): Implement "get_num_splits"
|
| 219 |
+
# There's an heuristic to compute num_splits when "num_splits <= 0"
|
| 220 |
+
# assert that num_splits is > 0 for now
|
| 221 |
+
if num_splits <= 0:
|
| 222 |
+
raise ValueError(f"tracing (torch.compile/torch.export) with num_splits <= 0 not supported. Got {num_splits=}")
|
| 223 |
+
|
| 224 |
+
if num_splits > 1:
|
| 225 |
+
if is_varlen_q:
|
| 226 |
+
out_accum = torch.empty((num_splits, num_heads, total_q, head_size_v), dtype=torch.float32, device=q.device)
|
| 227 |
+
softmax_lse_accum = torch.empty((num_splits, num_heads, total_q), dtype=torch.float32, device=q.device)
|
| 228 |
+
else:
|
| 229 |
+
out_accum = torch.empty((num_splits, batch_size, num_heads, seqlen_q, head_size_v), dtype=torch.float32, device=q.device)
|
| 230 |
+
softmax_lse_accum = torch.empty((num_splits, batch_size, num_heads, seqlen_q), dtype=torch.float32, device=q.device)
|
| 231 |
+
else:
|
| 232 |
+
# Tensors are not set when num_splits < 1
|
| 233 |
+
out_accum = torch.tensor([], device=out.device)
|
| 234 |
+
softmax_lse_accum = torch.tensor([], device=out.device)
|
| 235 |
+
|
| 236 |
+
return out, softmax_lse, out_accum, softmax_lse_accum
|
| 237 |
+
|
| 238 |
+
|
| 239 |
+
@torch.library.custom_op(add_op_namespace_prefix("_flash_attn_backward"), mutates_args=("dq", "dk", "dv"), device_types="cuda")
|
| 240 |
+
def _flash_attn_backward(
|
| 241 |
+
dout: torch.Tensor,
|
| 242 |
+
q: torch.Tensor,
|
| 243 |
+
k: torch.Tensor,
|
| 244 |
+
v: torch.Tensor,
|
| 245 |
+
out: torch.Tensor,
|
| 246 |
+
softmax_lse: torch.Tensor,
|
| 247 |
+
cu_seqlens_q: Optional[torch.Tensor] = None,
|
| 248 |
+
cu_seqlens_k: Optional[torch.Tensor] = None,
|
| 249 |
+
sequed_q: Optional[torch.Tensor] = None,
|
| 250 |
+
sequed_k: Optional[torch.Tensor] = None,
|
| 251 |
+
max_seqlen_q: Optional[int] = None,
|
| 252 |
+
max_seqlen_k: Optional[int] = None,
|
| 253 |
+
dq: Optional[torch.Tensor] = None,
|
| 254 |
+
dk: Optional[torch.Tensor] = None,
|
| 255 |
+
dv: Optional[torch.Tensor] = None,
|
| 256 |
+
softmax_scale: Optional[float] = None,
|
| 257 |
+
is_causal: bool = False,
|
| 258 |
+
window_size_left: int = -1,
|
| 259 |
+
window_size_right: int = -1,
|
| 260 |
+
softcap: float = 0.0,
|
| 261 |
+
deterministic: bool = False,
|
| 262 |
+
sm_margin: int = 0,
|
| 263 |
+
) -> torch.Tensor:
|
| 264 |
+
# dq, dk, dv are allocated by us so they should already be contiguous
|
| 265 |
+
dout, q, k, v, out = [maybe_contiguous(x) for x in (dout, q, k, v, out)]
|
| 266 |
+
softmax_d, *rest = flash_attn_3_cuda.bwd(
|
| 267 |
+
dout,
|
| 268 |
+
q,
|
| 269 |
+
k,
|
| 270 |
+
v,
|
| 271 |
+
out,
|
| 272 |
+
softmax_lse,
|
| 273 |
+
dq,
|
| 274 |
+
dk,
|
| 275 |
+
dv,
|
| 276 |
+
cu_seqlens_q,
|
| 277 |
+
cu_seqlens_k,
|
| 278 |
+
sequed_q,
|
| 279 |
+
sequed_k,
|
| 280 |
+
max_seqlen_q,
|
| 281 |
+
max_seqlen_k,
|
| 282 |
+
softmax_scale,
|
| 283 |
+
is_causal,
|
| 284 |
+
window_size_left,
|
| 285 |
+
window_size_right,
|
| 286 |
+
softcap,
|
| 287 |
+
deterministic,
|
| 288 |
+
sm_margin,
|
| 289 |
+
)
|
| 290 |
+
return softmax_d
|
| 291 |
+
|
| 292 |
+
|
| 293 |
+
@torch.library.register_fake(add_op_namespace_prefix("_flash_attn_backward"))
|
| 294 |
+
def _flash_attn_backward_fake(
|
| 295 |
+
dout: torch.Tensor,
|
| 296 |
+
q: torch.Tensor,
|
| 297 |
+
k: torch.Tensor,
|
| 298 |
+
v: torch.Tensor,
|
| 299 |
+
out: torch.Tensor,
|
| 300 |
+
softmax_lse: torch.Tensor,
|
| 301 |
+
cu_seqlens_q: Optional[torch.Tensor] = None,
|
| 302 |
+
cu_seqlens_k: Optional[torch.Tensor] = None,
|
| 303 |
+
sequed_q: Optional[torch.Tensor] = None,
|
| 304 |
+
sequed_k: Optional[torch.Tensor] = None,
|
| 305 |
+
max_seqlen_q: Optional[int] = None,
|
| 306 |
+
max_seqlen_k: Optional[int] = None,
|
| 307 |
+
dq: Optional[torch.Tensor] = None,
|
| 308 |
+
dk: Optional[torch.Tensor] = None,
|
| 309 |
+
dv: Optional[torch.Tensor] = None,
|
| 310 |
+
softmax_scale: Optional[float] = None,
|
| 311 |
+
is_causal: bool = False,
|
| 312 |
+
window_size_left: int = -1,
|
| 313 |
+
window_size_right: int = -1,
|
| 314 |
+
softcap: float = 0.0,
|
| 315 |
+
deterministic: bool = False,
|
| 316 |
+
sm_margin: int = 0,
|
| 317 |
+
) -> torch.Tensor:
|
| 318 |
+
|
| 319 |
+
is_varlen_q = cu_seqlens_q is not None
|
| 320 |
+
is_varlen_k = cu_seqlens_q is not None
|
| 321 |
+
is_varlen = is_varlen_q or is_varlen_k or sequed_q is not None or sequed_k is not None
|
| 322 |
+
|
| 323 |
+
if not is_varlen_q:
|
| 324 |
+
batch_size = q.size(0)
|
| 325 |
+
seqlen_q = q.size(1)
|
| 326 |
+
seqlen_k = k.size(1)
|
| 327 |
+
total_q = batch_size * q.size(1)
|
| 328 |
+
else:
|
| 329 |
+
batch_size = cu_seqlens_q.size(0) - 1
|
| 330 |
+
total_q = q.size(0)
|
| 331 |
+
seqlen_q = max_seqlen_q
|
| 332 |
+
seqlen_k = max_seqlen_k
|
| 333 |
+
|
| 334 |
+
if window_size_left >= seqlen_k - 1:
|
| 335 |
+
window_size_left = -1
|
| 336 |
+
|
| 337 |
+
if window_size_right >= seqlen_q - 1:
|
| 338 |
+
window_size_right = -1
|
| 339 |
+
|
| 340 |
+
if is_causal:
|
| 341 |
+
window_size_right = 0
|
| 342 |
+
|
| 343 |
+
is_causal = window_size_left < 0 and window_size_right == 0
|
| 344 |
+
|
| 345 |
+
head_size = q.size(-1)
|
| 346 |
+
head_size_v = v.size(-1)
|
| 347 |
+
head_size_rounded = round_up_headdim(max(head_size, head_size_v))
|
| 348 |
+
|
| 349 |
+
# Hopper gpus uses cuda compute capabilities 9.0
|
| 350 |
+
cap = torch.cuda.get_device_capability(q.device)
|
| 351 |
+
arch = cap[0] * 10 + cap[1]
|
| 352 |
+
|
| 353 |
+
is_local = (window_size_left >= 0 or window_size_right >= 0) and not is_causal
|
| 354 |
+
|
| 355 |
+
if head_size_rounded <= 64:
|
| 356 |
+
kBlockM_sm90 = 96 if (is_causal and softcap > 0.0) else 128
|
| 357 |
+
elif head_size_rounded <= 96:
|
| 358 |
+
kBlockM_sm90 = 64
|
| 359 |
+
elif head_size_rounded <= 128:
|
| 360 |
+
kBlockM_sm90 = 64 if (is_causal or is_local or softcap > 0.0) else 80
|
| 361 |
+
else:
|
| 362 |
+
kBlockM_sm90 = 64
|
| 363 |
+
|
| 364 |
+
kBlockM_sm80 = 128 if head_size_rounded <= 64 else 64
|
| 365 |
+
kBlockM_sm86 = 64 if head_size_rounded <= 192 else 32
|
| 366 |
+
|
| 367 |
+
if arch >= 90:
|
| 368 |
+
kBlockM = kBlockM_sm90
|
| 369 |
+
elif arch == 86 or arch == 89:
|
| 370 |
+
kBlockM = kBlockM_sm86
|
| 371 |
+
else:
|
| 372 |
+
kBlockM = kBlockM_sm80
|
| 373 |
+
|
| 374 |
+
num_heads = q.shape[-2]
|
| 375 |
+
seqlen_q_rounded = round_multiple(seqlen_q, kBlockM)
|
| 376 |
+
|
| 377 |
+
total_q_padded_rounded = round_multiple(total_q + batch_size * kBlockM, kBlockM)
|
| 378 |
+
|
| 379 |
+
dq = torch.empty_like(q) if dq is None else dq
|
| 380 |
+
dk = torch.empty_like(k) if dk is None else dk
|
| 381 |
+
dv = torch.empty_like(v) if dv is None else dv
|
| 382 |
+
|
| 383 |
+
if not is_varlen:
|
| 384 |
+
softmax_d = torch.empty((batch_size, num_heads, seqlen_q_rounded), dtype=torch.float32, device=q.device)
|
| 385 |
+
else:
|
| 386 |
+
softmax_d = torch.empty((num_heads, total_q_padded_rounded), dtype=torch.float32, device=q.device)
|
| 387 |
+
|
| 388 |
+
return softmax_d
|
| 389 |
+
|
| 390 |
+
|
| 391 |
+
def setup_context(ctx, inputs, output):
|
| 392 |
+
q, k, v = inputs[:3]
|
| 393 |
+
out, softmax_lse, _, _ = output
|
| 394 |
+
ctx.save_for_backward(q, k, v, out, softmax_lse)
|
| 395 |
+
ctx.softmax_scale = inputs[-11]
|
| 396 |
+
ctx.causal = inputs[-10]
|
| 397 |
+
ctx.window_size = [inputs[-9], inputs[-8]]
|
| 398 |
+
ctx.attention_chunk = inputs[-7]
|
| 399 |
+
ctx.softcap = inputs[-6]
|
| 400 |
+
ctx.sm_margin = inputs[-1]
|
| 401 |
+
|
| 402 |
+
|
| 403 |
+
def _backward(ctx, dout, *grads):
|
| 404 |
+
q, k, v, out, softmax_lse = ctx.saved_tensors
|
| 405 |
+
dq, dk, dv = torch.empty_like(q), torch.empty_like(k), torch.empty_like(v)
|
| 406 |
+
_flash_attn_backward(
|
| 407 |
+
dout,
|
| 408 |
+
q,
|
| 409 |
+
k,
|
| 410 |
+
v,
|
| 411 |
+
out,
|
| 412 |
+
softmax_lse,
|
| 413 |
+
None, None, # cu_seqlens_q, cu_seqlens_k,
|
| 414 |
+
None, None, # sequed_q, sequed_k,
|
| 415 |
+
None, None, # max_seqlen_q, max_seqlen_k,
|
| 416 |
+
dq,
|
| 417 |
+
dk,
|
| 418 |
+
dv,
|
| 419 |
+
ctx.softmax_scale,
|
| 420 |
+
ctx.causal,
|
| 421 |
+
ctx.window_size[0],
|
| 422 |
+
ctx.window_size[1],
|
| 423 |
+
ctx.softcap,
|
| 424 |
+
False, # deterministic
|
| 425 |
+
ctx.sm_margin,
|
| 426 |
+
)
|
| 427 |
+
return dq, dk, dv, *((None,) * 21)
|
| 428 |
+
|
| 429 |
+
|
| 430 |
+
_flash_attn_forward.register_autograd(_backward, setup_context=setup_context)
|
| 431 |
+
|
| 432 |
+
|
| 433 |
+
|
| 434 |
+
class FlashAttnQKVPackedFunc(torch.autograd.Function):
|
| 435 |
+
@staticmethod
|
| 436 |
+
def forward(
|
| 437 |
+
ctx,
|
| 438 |
+
qkv,
|
| 439 |
+
softmax_scale,
|
| 440 |
+
causal,
|
| 441 |
+
q_descale=None, k_descale=None, v_descale=None,
|
| 442 |
+
window_size=(-1, -1),
|
| 443 |
+
attention_chunk=0,
|
| 444 |
+
softcap=0.0,
|
| 445 |
+
deterministic=False,
|
| 446 |
+
num_heads_q=None,
|
| 447 |
+
sm_margin=0,
|
| 448 |
+
return_softmax=False,
|
| 449 |
+
):
|
| 450 |
+
if softmax_scale is None:
|
| 451 |
+
softmax_scale = qkv.shape[-1] ** (-0.5)
|
| 452 |
+
if qkv.dim() == 5:
|
| 453 |
+
assert qkv.shape[-3] == 3
|
| 454 |
+
q, k, v = qkv.unbind(dim=-3)
|
| 455 |
+
else:
|
| 456 |
+
assert qkv.dim() == 4
|
| 457 |
+
assert num_heads_q is not None
|
| 458 |
+
num_heads_k = (qkv.shape[2] - num_heads_q) // 2
|
| 459 |
+
assert num_heads_k * 2 + num_heads_q == qkv.shape[2]
|
| 460 |
+
q, k, v = qkv.split([num_heads_q, num_heads_k, num_heads_k], dim=-2)
|
| 461 |
+
out, softmax_lse, *rest = _flash_attn_forward(
|
| 462 |
+
q,
|
| 463 |
+
k,
|
| 464 |
+
v,
|
| 465 |
+
None, None, # k_new, v_new
|
| 466 |
+
None, # qv
|
| 467 |
+
None, # out
|
| 468 |
+
None, None, None, # cu_seqlens_q/k/k_new
|
| 469 |
+
None, None, # seqused_q/k
|
| 470 |
+
None, None, # max_seqlen_q/k
|
| 471 |
+
None, None, None, # page_table, kv_batch_idx, leftpad_k,
|
| 472 |
+
None, None, None, # rotary_cos/sin, seqlens_rotary
|
| 473 |
+
q_descale, k_descale, v_descale,
|
| 474 |
+
softmax_scale,
|
| 475 |
+
causal=causal,
|
| 476 |
+
window_size_left=window_size[0],
|
| 477 |
+
window_size_right=window_size[1],
|
| 478 |
+
attention_chunk=attention_chunk,
|
| 479 |
+
softcap=softcap,
|
| 480 |
+
sm_margin=sm_margin,
|
| 481 |
+
)
|
| 482 |
+
# ctx.save_for_backward(q, k, v, out_padded, softmax_lse)
|
| 483 |
+
ctx.save_for_backward(q, k, v, out, softmax_lse)
|
| 484 |
+
ctx.softmax_scale = softmax_scale
|
| 485 |
+
ctx.causal = causal
|
| 486 |
+
ctx.window_size = window_size
|
| 487 |
+
ctx.attention_chunk = attention_chunk
|
| 488 |
+
ctx.softcap = softcap
|
| 489 |
+
ctx.deterministic = deterministic
|
| 490 |
+
ctx.ndim = qkv.dim()
|
| 491 |
+
ctx.sm_margin = sm_margin
|
| 492 |
+
return (out, softmax_lse) if return_softmax else out
|
| 493 |
+
|
| 494 |
+
@staticmethod
|
| 495 |
+
def backward(ctx, dout, *args):
|
| 496 |
+
q, k, v, out, softmax_lse = ctx.saved_tensors
|
| 497 |
+
assert ctx.attention_chunk == 0, "FA3 backward does not support attention_chunk"
|
| 498 |
+
if ctx.ndim == 5:
|
| 499 |
+
qkv_shape = q.shape[:-2] + (3, *q.shape[-2:])
|
| 500 |
+
dqkv = torch.empty(qkv_shape, dtype=q.dtype, device=q.device)
|
| 501 |
+
dq, dk, dv = dqkv.unbind(dim=-3)
|
| 502 |
+
else:
|
| 503 |
+
num_heads_q = q.shape[2]
|
| 504 |
+
num_heads_k = k.shape[2]
|
| 505 |
+
qkv_shape = q.shape[:-2] + (num_heads_q + num_heads_k * 2, *q.shape[-1:])
|
| 506 |
+
dqkv = torch.empty(qkv_shape, dtype=q.dtype, device=q.device)
|
| 507 |
+
dq, dk, dv = dqkv.split([num_heads_q, num_heads_k, num_heads_k], dim=-2)
|
| 508 |
+
_flash_attn_backward(
|
| 509 |
+
dout,
|
| 510 |
+
q,
|
| 511 |
+
k,
|
| 512 |
+
v,
|
| 513 |
+
out,
|
| 514 |
+
softmax_lse,
|
| 515 |
+
None, None, # cu_seqlens_q, cu_seqlens_k,
|
| 516 |
+
None, None, # sequed_q, sequed_k,
|
| 517 |
+
None, None, # max_seqlen_q, max_seqlen_k,
|
| 518 |
+
dq,
|
| 519 |
+
dk,
|
| 520 |
+
dv,
|
| 521 |
+
ctx.softmax_scale,
|
| 522 |
+
ctx.causal,
|
| 523 |
+
ctx.window_size[0],
|
| 524 |
+
ctx.window_size[1],
|
| 525 |
+
ctx.softcap,
|
| 526 |
+
ctx.deterministic,
|
| 527 |
+
ctx.sm_margin,
|
| 528 |
+
)
|
| 529 |
+
dqkv = dqkv[..., : dout.shape[-1]] # We could have padded the head dimension
|
| 530 |
+
return dqkv, None, None, None, None, None, None, None, None, None, None, None, None
|
| 531 |
+
|
| 532 |
+
|
| 533 |
+
class FlashAttnFunc(torch.autograd.Function):
|
| 534 |
+
|
| 535 |
+
@staticmethod
|
| 536 |
+
def forward(
|
| 537 |
+
ctx,
|
| 538 |
+
q,
|
| 539 |
+
k,
|
| 540 |
+
v,
|
| 541 |
+
softmax_scale,
|
| 542 |
+
causal,
|
| 543 |
+
qv=None,
|
| 544 |
+
q_descale=None, k_descale=None, v_descale=None,
|
| 545 |
+
window_size=(-1, -1),
|
| 546 |
+
attention_chunk=0,
|
| 547 |
+
softcap=0.0,
|
| 548 |
+
num_splits=1,
|
| 549 |
+
pack_gqa=None,
|
| 550 |
+
deterministic=False,
|
| 551 |
+
sm_margin=0,
|
| 552 |
+
return_softmax=False,
|
| 553 |
+
):
|
| 554 |
+
if softmax_scale is None:
|
| 555 |
+
softmax_scale = (q.shape[-1] + (qv.shape[-1] if qv is not None else 0)) ** (-0.5)
|
| 556 |
+
# out, q, k, v, out_padded, softmax_lse = _flash_attn_forward(
|
| 557 |
+
out, softmax_lse, *rest = _flash_attn_forward(
|
| 558 |
+
q,
|
| 559 |
+
k,
|
| 560 |
+
v,
|
| 561 |
+
None, None, # k_new, v_new
|
| 562 |
+
qv, # qv
|
| 563 |
+
None, # out
|
| 564 |
+
None, None, None, # cu_seqlens_q/k/k_new
|
| 565 |
+
None, None, # seqused_q/k
|
| 566 |
+
None, None, # max_seqlen_q/k
|
| 567 |
+
None, None, None, # page_table, kv_batch_idx, leftpad_k,
|
| 568 |
+
None, None, None, # rotary_cos/sin, seqlens_rotary
|
| 569 |
+
q_descale, k_descale, v_descale,
|
| 570 |
+
softmax_scale,
|
| 571 |
+
causal=causal,
|
| 572 |
+
window_size_left=window_size[0],
|
| 573 |
+
window_size_right=window_size[1],
|
| 574 |
+
attention_chunk=attention_chunk,
|
| 575 |
+
softcap=softcap,
|
| 576 |
+
num_splits=num_splits,
|
| 577 |
+
pack_gqa=pack_gqa,
|
| 578 |
+
sm_margin=sm_margin,
|
| 579 |
+
)
|
| 580 |
+
# ctx.save_for_backward(q, k, v, out_padded, softmax_lse)
|
| 581 |
+
ctx.save_for_backward(q, k, v, out, softmax_lse)
|
| 582 |
+
ctx.softmax_scale = softmax_scale
|
| 583 |
+
ctx.causal = causal
|
| 584 |
+
ctx.window_size = window_size
|
| 585 |
+
ctx.attention_chunk = attention_chunk
|
| 586 |
+
ctx.softcap = softcap
|
| 587 |
+
ctx.deterministic = deterministic
|
| 588 |
+
ctx.sm_margin = sm_margin
|
| 589 |
+
return (out, softmax_lse) if return_softmax else out
|
| 590 |
+
|
| 591 |
+
@staticmethod
|
| 592 |
+
def backward(ctx, dout, *args):
|
| 593 |
+
q, k, v, out, softmax_lse = ctx.saved_tensors
|
| 594 |
+
assert ctx.attention_chunk == 0, "FA3 backward does not support attention_chunk"
|
| 595 |
+
dq, dk, dv = torch.empty_like(q), torch.empty_like(k), torch.empty_like(v)
|
| 596 |
+
_flash_attn_backward(
|
| 597 |
+
dout,
|
| 598 |
+
q,
|
| 599 |
+
k,
|
| 600 |
+
v,
|
| 601 |
+
out,
|
| 602 |
+
softmax_lse,
|
| 603 |
+
None, None, # cu_seqlens_q, cu_seqlens_k,
|
| 604 |
+
None, None, # sequed_q, sequed_k,
|
| 605 |
+
None, None, # max_seqlen_q, max_seqlen_k,
|
| 606 |
+
dq,
|
| 607 |
+
dk,
|
| 608 |
+
dv,
|
| 609 |
+
ctx.softmax_scale,
|
| 610 |
+
ctx.causal,
|
| 611 |
+
ctx.window_size[0],
|
| 612 |
+
ctx.window_size[1],
|
| 613 |
+
ctx.softcap,
|
| 614 |
+
ctx.deterministic,
|
| 615 |
+
ctx.sm_margin,
|
| 616 |
+
)
|
| 617 |
+
dq = dq[..., : q.shape[-1]] # We could have padded the head dimension
|
| 618 |
+
dk = dk[..., : k.shape[-1]]
|
| 619 |
+
dv = dv[..., : v.shape[-1]]
|
| 620 |
+
return dq, dk, dv, None, None, None, None, None, None, None, None, None, None, None, None, None, None
|
| 621 |
+
|
| 622 |
+
|
| 623 |
+
class FlashAttnVarlenFunc(torch.autograd.Function):
|
| 624 |
+
|
| 625 |
+
@staticmethod
|
| 626 |
+
def forward(
|
| 627 |
+
ctx,
|
| 628 |
+
q,
|
| 629 |
+
k,
|
| 630 |
+
v,
|
| 631 |
+
cu_seqlens_q,
|
| 632 |
+
cu_seqlens_k,
|
| 633 |
+
seqused_q,
|
| 634 |
+
seqused_k,
|
| 635 |
+
max_seqlen_q,
|
| 636 |
+
max_seqlen_k,
|
| 637 |
+
softmax_scale,
|
| 638 |
+
causal,
|
| 639 |
+
qv=None,
|
| 640 |
+
q_descale=None, k_descale=None, v_descale=None,
|
| 641 |
+
window_size=(-1, -1),
|
| 642 |
+
attention_chunk=0,
|
| 643 |
+
softcap=0.0,
|
| 644 |
+
num_splits=1,
|
| 645 |
+
pack_gqa=None,
|
| 646 |
+
deterministic=False,
|
| 647 |
+
sm_margin=0,
|
| 648 |
+
return_softmax=False,
|
| 649 |
+
):
|
| 650 |
+
if softmax_scale is None:
|
| 651 |
+
softmax_scale = (q.shape[-1] + (qv.shape[-1] if qv is not None else 0)) ** (-0.5)
|
| 652 |
+
# out, q, k, v, out_padded, softmax_lse = _flash_attn_varlen_forward(
|
| 653 |
+
out, softmax_lse, *rest = _flash_attn_forward(
|
| 654 |
+
q,
|
| 655 |
+
k,
|
| 656 |
+
v,
|
| 657 |
+
None, None, # k_new, v_new
|
| 658 |
+
qv, # qv
|
| 659 |
+
None, # out
|
| 660 |
+
cu_seqlens_q,
|
| 661 |
+
cu_seqlens_k,
|
| 662 |
+
None, # cu_seqlens_k_new
|
| 663 |
+
seqused_q,
|
| 664 |
+
seqused_k,
|
| 665 |
+
max_seqlen_q,
|
| 666 |
+
max_seqlen_k,
|
| 667 |
+
None, None, None, # page_table, kv_batch_idx, leftpad_k,
|
| 668 |
+
None, None, None, # rotary_cos/sin, seqlens_rotary
|
| 669 |
+
q_descale, k_descale, v_descale,
|
| 670 |
+
softmax_scale,
|
| 671 |
+
causal=causal,
|
| 672 |
+
window_size_left=window_size[0],
|
| 673 |
+
window_size_right=window_size[1],
|
| 674 |
+
attention_chunk=attention_chunk,
|
| 675 |
+
softcap=softcap,
|
| 676 |
+
num_splits=num_splits,
|
| 677 |
+
pack_gqa=pack_gqa,
|
| 678 |
+
sm_margin=sm_margin,
|
| 679 |
+
)
|
| 680 |
+
# ctx.save_for_backward(q, k, v, out_padded, softmax_lse, cu_seqlens_q, cu_seqlens_k, seqused_q, seqused_k)
|
| 681 |
+
ctx.save_for_backward(q, k, v, out, softmax_lse, cu_seqlens_q, cu_seqlens_k, seqused_q, seqused_k)
|
| 682 |
+
ctx.max_seqlen_q = max_seqlen_q
|
| 683 |
+
ctx.max_seqlen_k = max_seqlen_k
|
| 684 |
+
ctx.softmax_scale = softmax_scale
|
| 685 |
+
ctx.causal = causal
|
| 686 |
+
ctx.window_size = window_size
|
| 687 |
+
ctx.attention_chunk = attention_chunk
|
| 688 |
+
ctx.softcap = softcap
|
| 689 |
+
ctx.deterministic = deterministic
|
| 690 |
+
ctx.sm_margin = sm_margin
|
| 691 |
+
return (out, softmax_lse) if return_softmax else out
|
| 692 |
+
|
| 693 |
+
@staticmethod
|
| 694 |
+
def backward(ctx, dout, *args):
|
| 695 |
+
q, k, v, out, softmax_lse, cu_seqlens_q, cu_seqlens_k, seqused_q, seqused_k = ctx.saved_tensors
|
| 696 |
+
assert ctx.attention_chunk == 0, "FA3 backward does not support attention_chunk"
|
| 697 |
+
dq, dk, dv = torch.empty_like(q), torch.empty_like(k), torch.empty_like(v)
|
| 698 |
+
_flash_attn_backward(
|
| 699 |
+
dout,
|
| 700 |
+
q,
|
| 701 |
+
k,
|
| 702 |
+
v,
|
| 703 |
+
out,
|
| 704 |
+
softmax_lse,
|
| 705 |
+
cu_seqlens_q,
|
| 706 |
+
cu_seqlens_k,
|
| 707 |
+
seqused_q,
|
| 708 |
+
seqused_k,
|
| 709 |
+
ctx.max_seqlen_q,
|
| 710 |
+
ctx.max_seqlen_k,
|
| 711 |
+
dq,
|
| 712 |
+
dk,
|
| 713 |
+
dv,
|
| 714 |
+
ctx.softmax_scale,
|
| 715 |
+
ctx.causal,
|
| 716 |
+
ctx.window_size[0],
|
| 717 |
+
ctx.window_size[1],
|
| 718 |
+
ctx.softcap,
|
| 719 |
+
ctx.deterministic,
|
| 720 |
+
ctx.sm_margin,
|
| 721 |
+
)
|
| 722 |
+
dq = dq[..., : q.shape[-1]] # We could have padded the head dimension
|
| 723 |
+
dk = dk[..., : k.shape[-1]]
|
| 724 |
+
dv = dv[..., : v.shape[-1]]
|
| 725 |
+
return dq, dk, dv, None, None, None, None, None, None, None, None, None, None, None, None, None, None, None, None, None, None, None, None
|
| 726 |
+
|
| 727 |
+
|
| 728 |
+
def flash_attn_qkvpacked_func(
|
| 729 |
+
qkv,
|
| 730 |
+
softmax_scale=None,
|
| 731 |
+
causal=False,
|
| 732 |
+
q_descale=None, k_descale=None, v_descale=None,
|
| 733 |
+
window_size=(-1, -1),
|
| 734 |
+
attention_chunk=0,
|
| 735 |
+
softcap=0.0,
|
| 736 |
+
deterministic=False,
|
| 737 |
+
num_heads_q=None,
|
| 738 |
+
sm_margin=0,
|
| 739 |
+
return_attn_probs=False,
|
| 740 |
+
):
|
| 741 |
+
"""dropout_p should be set to 0.0 during evaluation
|
| 742 |
+
If Q, K, V are already stacked into 1 tensor, this function will be faster than
|
| 743 |
+
calling flash_attn_func on Q, K, V since the backward pass avoids explicit concatenation
|
| 744 |
+
of the gradients of Q, K, V.
|
| 745 |
+
For multi-query and grouped-query attention (MQA/GQA), please see
|
| 746 |
+
flash_attn_kvpacked_func and flash_attn_func.
|
| 747 |
+
|
| 748 |
+
If window_size != (-1, -1), implements sliding window local attention. Query at position i
|
| 749 |
+
will only attend to keys between [i - window_size[0], i + window_size[1]] inclusive.
|
| 750 |
+
|
| 751 |
+
Arguments:
|
| 752 |
+
qkv: (batch_size, seqlen, 3, nheads, headdim)
|
| 753 |
+
dropout_p: float. Dropout probability.
|
| 754 |
+
softmax_scale: float. The scaling of QK^T before applying softmax.
|
| 755 |
+
Default to 1 / sqrt(headdim).
|
| 756 |
+
causal: bool. Whether to apply causal attention mask (e.g., for auto-regressive modeling).
|
| 757 |
+
window_size: (left, right). If not (-1, -1), implements sliding window local attention.
|
| 758 |
+
softcap: float. Anything > 0 activates softcapping attention.
|
| 759 |
+
alibi_slopes: (nheads,) or (batch_size, nheads), fp32. A bias of (-alibi_slope * |i - j|) is added to
|
| 760 |
+
the attention score of query i and key j.
|
| 761 |
+
deterministic: bool. Whether to use the deterministic implementation of the backward pass,
|
| 762 |
+
which is slightly slower and uses more memory. The forward pass is always deterministic.
|
| 763 |
+
return_attn_probs: bool. Whether to return the attention probabilities. This option is for
|
| 764 |
+
testing only. The returned probabilities are not guaranteed to be correct
|
| 765 |
+
(they might not have the right scaling).
|
| 766 |
+
Return:
|
| 767 |
+
out: (batch_size, seqlen, nheads, headdim).
|
| 768 |
+
softmax_lse [optional, if return_attn_probs=True]: (batch_size, nheads, seqlen). The
|
| 769 |
+
logsumexp of each row of the matrix QK^T * scaling (e.g., log of the softmax
|
| 770 |
+
normalization factor).
|
| 771 |
+
S_dmask [optional, if return_attn_probs=True]: (batch_size, nheads, seqlen, seqlen).
|
| 772 |
+
The output of softmax (possibly with different scaling). It also encodes the dropout
|
| 773 |
+
pattern (negative means that location was dropped, nonnegative means it was kept).
|
| 774 |
+
"""
|
| 775 |
+
return FlashAttnQKVPackedFunc.apply(
|
| 776 |
+
qkv,
|
| 777 |
+
softmax_scale,
|
| 778 |
+
causal,
|
| 779 |
+
q_descale, k_descale, v_descale,
|
| 780 |
+
window_size,
|
| 781 |
+
attention_chunk,
|
| 782 |
+
softcap,
|
| 783 |
+
deterministic,
|
| 784 |
+
num_heads_q,
|
| 785 |
+
sm_margin,
|
| 786 |
+
return_attn_probs,
|
| 787 |
+
)
|
| 788 |
+
|
| 789 |
+
|
| 790 |
+
def flash_attn_func(
|
| 791 |
+
q,
|
| 792 |
+
k,
|
| 793 |
+
v,
|
| 794 |
+
softmax_scale=None,
|
| 795 |
+
causal=False,
|
| 796 |
+
qv=None,
|
| 797 |
+
q_descale=None, k_descale=None, v_descale=None,
|
| 798 |
+
window_size=(-1, -1),
|
| 799 |
+
attention_chunk=0,
|
| 800 |
+
softcap=0.0,
|
| 801 |
+
num_splits=1,
|
| 802 |
+
pack_gqa=None,
|
| 803 |
+
deterministic=False,
|
| 804 |
+
sm_margin=0,
|
| 805 |
+
return_attn_probs=False,
|
| 806 |
+
):
|
| 807 |
+
"""dropout_p should be set to 0.0 during evaluation
|
| 808 |
+
Supports multi-query and grouped-query attention (MQA/GQA) by passing in KV with fewer heads
|
| 809 |
+
than Q. Note that the number of heads in Q must be divisible by the number of heads in KV.
|
| 810 |
+
For example, if Q has 6 heads and K, V have 2 heads, head 0, 1, 2 of Q will attention to head
|
| 811 |
+
0 of K, V, and head 3, 4, 5 of Q will attention to head 1 of K, V.
|
| 812 |
+
|
| 813 |
+
If causal=True, the causal mask is aligned to the bottom right corner of the attention matrix.
|
| 814 |
+
For example, if seqlen_q = 2 and seqlen_k = 5, the causal mask (1 = keep, 0 = masked out) is:
|
| 815 |
+
1 1 1 1 0
|
| 816 |
+
1 1 1 1 1
|
| 817 |
+
If seqlen_q = 5 and seqlen_k = 2, the causal mask is:
|
| 818 |
+
0 0
|
| 819 |
+
0 0
|
| 820 |
+
0 0
|
| 821 |
+
1 0
|
| 822 |
+
1 1
|
| 823 |
+
If the row of the mask is all zero, the output will be zero.
|
| 824 |
+
|
| 825 |
+
If window_size != (-1, -1), implements sliding window local attention. Query at position i
|
| 826 |
+
will only attend to keys between
|
| 827 |
+
[i + seqlen_k - seqlen_q - window_size[0], i + seqlen_k - seqlen_q + window_size[1]] inclusive.
|
| 828 |
+
|
| 829 |
+
Arguments:
|
| 830 |
+
q: (batch_size, seqlen, nheads, headdim)
|
| 831 |
+
k: (batch_size, seqlen, nheads_k, headdim)
|
| 832 |
+
v: (batch_size, seqlen, nheads_k, headdim)
|
| 833 |
+
dropout_p: float. Dropout probability.
|
| 834 |
+
softmax_scale: float. The scaling of QK^T before applying softmax.
|
| 835 |
+
Default to 1 / sqrt(headdim).
|
| 836 |
+
causal: bool. Whether to apply causal attention mask (e.g., for auto-regressive modeling).
|
| 837 |
+
window_size: (left, right). If not (-1, -1), implements sliding window local attention.
|
| 838 |
+
alibi_slopes: (nheads,) or (batch_size, nheads), fp32. A bias of
|
| 839 |
+
(-alibi_slope * |i + seqlen_k - seqlen_q - j|)
|
| 840 |
+
is added to the attention score of query i and key j.
|
| 841 |
+
deterministic: bool. Whether to use the deterministic implementation of the backward pass,
|
| 842 |
+
which is slightly slower and uses more memory. The forward pass is always deterministic.
|
| 843 |
+
return_attn_probs: bool. Whether to return the attention probabilities. This option is for
|
| 844 |
+
testing only. The returned probabilities are not guaranteed to be correct
|
| 845 |
+
(they might not have the right scaling).
|
| 846 |
+
Return:
|
| 847 |
+
out: (batch_size, seqlen, nheads, headdim).
|
| 848 |
+
softmax_lse [optional, if return_attn_probs=True]: (batch_size, nheads, seqlen). The
|
| 849 |
+
logsumexp of each row of the matrix QK^T * scaling (e.g., log of the softmax
|
| 850 |
+
normalization factor).
|
| 851 |
+
"""
|
| 852 |
+
return FlashAttnFunc.apply(
|
| 853 |
+
q,
|
| 854 |
+
k,
|
| 855 |
+
v,
|
| 856 |
+
softmax_scale,
|
| 857 |
+
causal,
|
| 858 |
+
qv,
|
| 859 |
+
q_descale, k_descale, v_descale,
|
| 860 |
+
window_size,
|
| 861 |
+
attention_chunk,
|
| 862 |
+
softcap,
|
| 863 |
+
num_splits,
|
| 864 |
+
pack_gqa,
|
| 865 |
+
deterministic,
|
| 866 |
+
sm_margin,
|
| 867 |
+
return_attn_probs,
|
| 868 |
+
)
|
| 869 |
+
|
| 870 |
+
|
| 871 |
+
def flash_attn_varlen_func(
|
| 872 |
+
q,
|
| 873 |
+
k,
|
| 874 |
+
v,
|
| 875 |
+
cu_seqlens_q,
|
| 876 |
+
cu_seqlens_k,
|
| 877 |
+
max_seqlen_q,
|
| 878 |
+
max_seqlen_k,
|
| 879 |
+
seqused_q=None,
|
| 880 |
+
seqused_k=None,
|
| 881 |
+
softmax_scale=None,
|
| 882 |
+
causal=False,
|
| 883 |
+
qv=None,
|
| 884 |
+
q_descale=None, k_descale=None, v_descale=None,
|
| 885 |
+
window_size=(-1, -1),
|
| 886 |
+
attention_chunk=0,
|
| 887 |
+
softcap=0.0,
|
| 888 |
+
num_splits=1,
|
| 889 |
+
pack_gqa=None,
|
| 890 |
+
deterministic=False,
|
| 891 |
+
sm_margin=0,
|
| 892 |
+
return_attn_probs=False,
|
| 893 |
+
):
|
| 894 |
+
return FlashAttnVarlenFunc.apply(
|
| 895 |
+
q,
|
| 896 |
+
k,
|
| 897 |
+
v,
|
| 898 |
+
cu_seqlens_q,
|
| 899 |
+
cu_seqlens_k,
|
| 900 |
+
seqused_q,
|
| 901 |
+
seqused_k,
|
| 902 |
+
max_seqlen_q,
|
| 903 |
+
max_seqlen_k,
|
| 904 |
+
softmax_scale,
|
| 905 |
+
causal,
|
| 906 |
+
qv,
|
| 907 |
+
q_descale, k_descale, v_descale,
|
| 908 |
+
window_size,
|
| 909 |
+
attention_chunk,
|
| 910 |
+
softcap,
|
| 911 |
+
num_splits,
|
| 912 |
+
pack_gqa,
|
| 913 |
+
deterministic,
|
| 914 |
+
sm_margin,
|
| 915 |
+
return_attn_probs,
|
| 916 |
+
)
|
| 917 |
+
|
| 918 |
+
|
| 919 |
+
def flash_attn_combine(out_partial, lse_partial, out=None, out_dtype=None):
|
| 920 |
+
return flash_attn_3_cuda.fwd_combine(out_partial, lse_partial, out, out_dtype)
|
| 921 |
+
|
| 922 |
+
|
| 923 |
+
def flash_attn_with_kvcache(
|
| 924 |
+
q,
|
| 925 |
+
k_cache,
|
| 926 |
+
v_cache,
|
| 927 |
+
k=None,
|
| 928 |
+
v=None,
|
| 929 |
+
qv=None,
|
| 930 |
+
rotary_cos=None,
|
| 931 |
+
rotary_sin=None,
|
| 932 |
+
cache_seqlens: Optional[Union[(int, torch.Tensor)]] = None,
|
| 933 |
+
cache_batch_idx: Optional[torch.Tensor] = None,
|
| 934 |
+
cache_leftpad: Optional[torch.Tensor] = None,
|
| 935 |
+
page_table: Optional[torch.Tensor] = None,
|
| 936 |
+
cu_seqlens_q: Optional[torch.Tensor] = None,
|
| 937 |
+
cu_seqlens_k_new: Optional[torch.Tensor] = None,
|
| 938 |
+
max_seqlen_q: Optional[int] = None,
|
| 939 |
+
rotary_seqlens: Optional[torch.Tensor] = None,
|
| 940 |
+
q_descale: Optional[torch.Tensor] = None,
|
| 941 |
+
k_descale: Optional[torch.Tensor] = None,
|
| 942 |
+
v_descale: Optional[torch.Tensor] = None,
|
| 943 |
+
softmax_scale=None,
|
| 944 |
+
causal=False,
|
| 945 |
+
window_size=(-1, -1), # -1 means infinite context window
|
| 946 |
+
attention_chunk=0,
|
| 947 |
+
softcap=0.0, # 0.0 means deactivated
|
| 948 |
+
rotary_interleaved=True,
|
| 949 |
+
scheduler_metadata=None,
|
| 950 |
+
num_splits=0, # Can be tuned for speed
|
| 951 |
+
pack_gqa=None, # Can be tuned for speed
|
| 952 |
+
sm_margin=0, # Can be tuned if some SMs are used for communication
|
| 953 |
+
return_softmax_lse=False,
|
| 954 |
+
):
|
| 955 |
+
"""
|
| 956 |
+
If k and v are not None, k_cache and v_cache will be updated *inplace* with the new values from
|
| 957 |
+
k and v. This is useful for incremental decoding: you can pass in the cached keys/values from
|
| 958 |
+
the previous step, and update them with the new keys/values from the current step, and do
|
| 959 |
+
attention with the updated cache, all in 1 kernel.
|
| 960 |
+
|
| 961 |
+
If you pass in k / v, you must make sure that the cache is large enough to hold the new values.
|
| 962 |
+
For example, the KV cache could be pre-allocated with the max sequence length, and you can use
|
| 963 |
+
cache_seqlens to keep track of the current sequence lengths of each sequence in the batch.
|
| 964 |
+
|
| 965 |
+
Also apply rotary embedding if rotary_cos and rotary_sin are passed in. The key @k will be
|
| 966 |
+
rotated by rotary_cos and rotary_sin at indices cache_seqlens, cache_seqlens + 1, etc.
|
| 967 |
+
If causal or local (i.e., window_size != (-1, -1)), the query @q will be rotated by rotary_cos
|
| 968 |
+
and rotary_sin at indices cache_seqlens, cache_seqlens + 1, etc.
|
| 969 |
+
If not causal and not local, the query @q will be rotated by rotary_cos and rotary_sin at
|
| 970 |
+
indices cache_seqlens only (i.e. we consider all tokens in @q to be at position cache_seqlens).
|
| 971 |
+
|
| 972 |
+
See tests/test_flash_attn.py::test_flash_attn_kvcache for examples of how to use this function.
|
| 973 |
+
|
| 974 |
+
Supports multi-query and grouped-query attention (MQA/GQA) by passing in KV with fewer heads
|
| 975 |
+
than Q. Note that the number of heads in Q must be divisible by the number of heads in KV.
|
| 976 |
+
For example, if Q has 6 heads and K, V have 2 heads, head 0, 1, 2 of Q will attention to head
|
| 977 |
+
0 of K, V, and head 3, 4, 5 of Q will attention to head 1 of K, V.
|
| 978 |
+
|
| 979 |
+
If causal=True, the causal mask is aligned to the bottom right corner of the attention matrix.
|
| 980 |
+
For example, if seqlen_q = 2 and seqlen_k = 5, the causal mask (1 = keep, 0 = masked out) is:
|
| 981 |
+
1 1 1 1 0
|
| 982 |
+
1 1 1 1 1
|
| 983 |
+
If seqlen_q = 5 and seqlen_k = 2, the causal mask is:
|
| 984 |
+
0 0
|
| 985 |
+
0 0
|
| 986 |
+
0 0
|
| 987 |
+
1 0
|
| 988 |
+
1 1
|
| 989 |
+
If the row of the mask is all zero, the output will be zero.
|
| 990 |
+
|
| 991 |
+
If window_size != (-1, -1), implements sliding window local attention. Query at position i
|
| 992 |
+
will only attend to keys between
|
| 993 |
+
[i + seqlen_k - seqlen_q - window_size[0], i + seqlen_k - seqlen_q + window_size[1]] inclusive.
|
| 994 |
+
|
| 995 |
+
Note: Does not support backward pass.
|
| 996 |
+
|
| 997 |
+
Arguments:
|
| 998 |
+
q: (batch_size, seqlen, nheads, headdim)
|
| 999 |
+
k_cache: (batch_size_cache, seqlen_cache, nheads_k, headdim) if there's no page_table,
|
| 1000 |
+
or (num_blocks, page_block_size, nheads_k, headdim) if there's a page_table (i.e. paged KV cache)
|
| 1001 |
+
page_block_size can be arbitrary (e.g, 1, 2, 3, 64, etc.).
|
| 1002 |
+
v_cache: (batch_size_cache, seqlen_cache, nheads_k, headdim_v) if there's no page_table,
|
| 1003 |
+
or (num_blocks, page_block_size, nheads_k, headdim_v) if there's a page_table (i.e. paged KV cache)
|
| 1004 |
+
k [optional]: (batch_size, seqlen_new, nheads_k, headdim). If not None, we concatenate
|
| 1005 |
+
k with k_cache, starting at the indices specified by cache_seqlens.
|
| 1006 |
+
v [optional]: (batch_size, seqlen_new, nheads_k, headdim_v). Similar to k.
|
| 1007 |
+
qv [optional]: (batch_size, seqlen, nheads, headdim_v)
|
| 1008 |
+
rotary_cos [optional]: (seqlen_ro, rotary_dim / 2). If not None, we apply rotary embedding
|
| 1009 |
+
to k and q. Only applicable if k and v are passed in. rotary_dim must be divisible by 16.
|
| 1010 |
+
rotary_sin [optional]: (seqlen_ro, rotary_dim / 2). Similar to rotary_cos.
|
| 1011 |
+
cache_seqlens: int, or (batch_size,), dtype torch.int32. The sequence lengths of the
|
| 1012 |
+
KV cache.
|
| 1013 |
+
cache_batch_idx: (batch_size,), dtype torch.int32. The indices used to index into the KV cache.
|
| 1014 |
+
If None, we assume that the batch indices are [0, 1, 2, ..., batch_size - 1].
|
| 1015 |
+
If the indices are not distinct, and k and v are provided, the values updated in the cache
|
| 1016 |
+
might come from any of the duplicate indices.
|
| 1017 |
+
cache_leftpad: (batch_size,), dtype torch.int32. The index that the KV cache starts. If None, assume 0.
|
| 1018 |
+
page_table [optional]: (batch_size, max_num_blocks_per_seq), dtype torch.int32.
|
| 1019 |
+
softmax_scale: float. The scaling of QK^T before applying softmax.
|
| 1020 |
+
Default to 1 / sqrt(headdim).
|
| 1021 |
+
causal: bool. Whether to apply causal attention mask (e.g., for auto-regressive modeling).
|
| 1022 |
+
window_size: (left, right). If not (-1, -1), implements sliding window local attention.
|
| 1023 |
+
softcap: float. Anything > 0 activates softcapping attention.
|
| 1024 |
+
rotary_interleaved: bool. Only applicable if rotary_cos and rotary_sin are passed in.
|
| 1025 |
+
If True, rotary embedding will combine dimensions 0 & 1, 2 & 3, etc. If False,
|
| 1026 |
+
rotary embedding will combine dimensions 0 & rotary_dim / 2, 1 & rotary_dim / 2 + 1
|
| 1027 |
+
(i.e. GPT-NeoX style).
|
| 1028 |
+
num_splits: int. If > 1, split the key/value into this many chunks along the sequence.
|
| 1029 |
+
If num_splits == 1, we don't split the key/value. If num_splits == 0, we use a heuristic
|
| 1030 |
+
to automatically determine the number of splits.
|
| 1031 |
+
Don't change this unless you know what you are doing.
|
| 1032 |
+
return_softmax_lse: bool. Whether to return the logsumexp of the attention scores.
|
| 1033 |
+
|
| 1034 |
+
Return:
|
| 1035 |
+
out: (batch_size, seqlen, nheads, headdim).
|
| 1036 |
+
softmax_lse [optional, if return_softmax_lse=True]: (batch_size, nheads, seqlen). The
|
| 1037 |
+
logsumexp of each row of the matrix QK^T * scaling (e.g., log of the softmax
|
| 1038 |
+
normalization factor).
|
| 1039 |
+
"""
|
| 1040 |
+
assert k_cache.stride(-1) == 1, "k_cache must have contiguous last dimension"
|
| 1041 |
+
assert v_cache.stride(-1) == 1, "v_cache must have contiguous last dimension"
|
| 1042 |
+
if softmax_scale is None:
|
| 1043 |
+
softmax_scale = (q.shape[-1] + (qv.shape[-1] if qv is not None else 0)) ** (-0.5)
|
| 1044 |
+
if cache_seqlens is not None and isinstance(cache_seqlens, int):
|
| 1045 |
+
cache_seqlens = torch.full(
|
| 1046 |
+
(q.shape[0],), cache_seqlens, dtype=torch.int32, device=k_cache.device
|
| 1047 |
+
)
|
| 1048 |
+
cache_seqlens = maybe_contiguous(cache_seqlens)
|
| 1049 |
+
out, softmax_lse, *rest = _flash_attn_forward(
|
| 1050 |
+
q,
|
| 1051 |
+
k_cache,
|
| 1052 |
+
v_cache,
|
| 1053 |
+
k,
|
| 1054 |
+
v,
|
| 1055 |
+
qv,
|
| 1056 |
+
None, # out
|
| 1057 |
+
cu_seqlens_q,
|
| 1058 |
+
None, # cu_seqlens_k
|
| 1059 |
+
cu_seqlens_k_new,
|
| 1060 |
+
None, # seqused_q
|
| 1061 |
+
cache_seqlens,
|
| 1062 |
+
max_seqlen_q,
|
| 1063 |
+
None, # max_seqlen_k
|
| 1064 |
+
page_table,
|
| 1065 |
+
cache_batch_idx,
|
| 1066 |
+
cache_leftpad,
|
| 1067 |
+
rotary_cos,
|
| 1068 |
+
rotary_sin,
|
| 1069 |
+
rotary_seqlens,
|
| 1070 |
+
q_descale, k_descale, v_descale,
|
| 1071 |
+
softmax_scale,
|
| 1072 |
+
causal=causal,
|
| 1073 |
+
window_size_left=window_size[0],
|
| 1074 |
+
window_size_right=window_size[1],
|
| 1075 |
+
attention_chunk=attention_chunk,
|
| 1076 |
+
softcap=softcap,
|
| 1077 |
+
rotary_interleaved=rotary_interleaved,
|
| 1078 |
+
scheduler_metadata=scheduler_metadata,
|
| 1079 |
+
num_splits=num_splits,
|
| 1080 |
+
pack_gqa=pack_gqa,
|
| 1081 |
+
sm_margin=sm_margin,
|
| 1082 |
+
)
|
| 1083 |
+
# return (out, softmax_lse) if return_softmax_lse else out
|
| 1084 |
+
return (out, softmax_lse, *rest) if return_softmax_lse else out
|
| 1085 |
+
|
| 1086 |
+
|
| 1087 |
+
def get_scheduler_metadata(
|
| 1088 |
+
batch_size, max_seqlen_q, max_seqlen_k, num_heads_q, num_heads_kv, headdim,
|
| 1089 |
+
cache_seqlens: torch.Tensor,
|
| 1090 |
+
qkv_dtype=torch.bfloat16,
|
| 1091 |
+
headdim_v=None,
|
| 1092 |
+
cu_seqlens_q: Optional[torch.Tensor] = None,
|
| 1093 |
+
cu_seqlens_k_new: Optional[torch.Tensor] = None,
|
| 1094 |
+
cache_leftpad: Optional[torch.Tensor] = None,
|
| 1095 |
+
page_size: Optional[int] = None,
|
| 1096 |
+
max_seqlen_k_new=0,
|
| 1097 |
+
causal=False,
|
| 1098 |
+
window_size=(-1, -1), # -1 means infinite context window
|
| 1099 |
+
attention_chunk=0,
|
| 1100 |
+
has_softcap=False,
|
| 1101 |
+
num_splits=0, # Can be tuned for speed
|
| 1102 |
+
pack_gqa=None, # Can be tuned for speed
|
| 1103 |
+
sm_margin=0, # Can be tuned if some SMs are used for communication
|
| 1104 |
+
):
|
| 1105 |
+
cache_seqlens = maybe_contiguous(cache_seqlens)
|
| 1106 |
+
if headdim_v is None:
|
| 1107 |
+
headdim_v = headdim
|
| 1108 |
+
scheduler_metadata = flash_attn_3_cuda.get_scheduler_metadata(
|
| 1109 |
+
batch_size, max_seqlen_q, max_seqlen_k, num_heads_q, num_heads_kv, headdim, headdim_v,
|
| 1110 |
+
qkv_dtype,
|
| 1111 |
+
cache_seqlens,
|
| 1112 |
+
cu_seqlens_q,
|
| 1113 |
+
None, # cu_seqlens_k
|
| 1114 |
+
cu_seqlens_k_new,
|
| 1115 |
+
None, # seqused_q
|
| 1116 |
+
cache_leftpad,
|
| 1117 |
+
page_size,
|
| 1118 |
+
max_seqlen_k_new,
|
| 1119 |
+
causal,
|
| 1120 |
+
window_size[0], window_size[1],
|
| 1121 |
+
attention_chunk,
|
| 1122 |
+
has_softcap,
|
| 1123 |
+
num_splits,
|
| 1124 |
+
pack_gqa,
|
| 1125 |
+
sm_margin,
|
| 1126 |
+
)
|
| 1127 |
+
return scheduler_metadata
|
build/torch-stable-abi29-cu128-x86_64-linux/metadata.json
ADDED
|
@@ -0,0 +1,38 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"name": "flash-attn3",
|
| 3 |
+
"id": "_flash_attn3_cuda_fe2bb66",
|
| 4 |
+
"version": 2,
|
| 5 |
+
"kernels-minver": "0.14.0",
|
| 6 |
+
"license": "BSD-3-Clause",
|
| 7 |
+
"python-depends": [],
|
| 8 |
+
"kernel-depends": [],
|
| 9 |
+
"backend": {
|
| 10 |
+
"type": "cuda",
|
| 11 |
+
"archs": [
|
| 12 |
+
"8.0",
|
| 13 |
+
"9.0a"
|
| 14 |
+
]
|
| 15 |
+
},
|
| 16 |
+
"digest": {
|
| 17 |
+
"algorithm": "sha256",
|
| 18 |
+
"files": {
|
| 19 |
+
"__init__.py": "b/45TnpaDDuTwp84ILd3Q+pQCIk2RnHtSSZKRq9Lj5E=",
|
| 20 |
+
"_flash_attn3_cuda_fe2bb66.abi3.so": "G7UnxhaXDrMLIulCS/ceSoiaAU8xgVlTH5r089qnfJQ=",
|
| 21 |
+
"_ops.py": "qE76YSn3lYonVMHeyk13xZljOBdqQRapvPt15D/lvSI=",
|
| 22 |
+
"flash_attn_config.py": "uxo+eyMcDit//8YDaS4Rtk5BTMKzTSK9tOKTjdWVe/w=",
|
| 23 |
+
"flash_attn_interface.py": "y0bD2JYGMFgX6iDre3zJN2sBPuq/12N6Mo55Z+HQCNQ=",
|
| 24 |
+
"symbols.json": "dGYI9GazLj56q85OhdJTeGBnq6b9RzaJ911NQsrsjhs="
|
| 25 |
+
}
|
| 26 |
+
},
|
| 27 |
+
"provenance": {
|
| 28 |
+
"kernel-builder": {
|
| 29 |
+
"version": "0.18.0-dev0",
|
| 30 |
+
"commit": "d054906a13710b811b96697b4635f03020842733",
|
| 31 |
+
"dirty": false
|
| 32 |
+
},
|
| 33 |
+
"kernel": {
|
| 34 |
+
"commit": "fe2bb66fee6c2a77eccf08a0508dc57428d488ed",
|
| 35 |
+
"dirty": false
|
| 36 |
+
}
|
| 37 |
+
}
|
| 38 |
+
}
|
build/torch-stable-abi29-cu128-x86_64-linux/metadata.json.sigstore
ADDED
|
@@ -0,0 +1 @@
|
|
|
|
|
|
|
| 1 |
+
{"mediaType":"application/vnd.dev.sigstore.bundle.v0.3+json", "verificationMaterial":{"certificate":{"rawBytes":"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"}, "tlogEntries":[{"logIndex":"3322704864", "logId":{"keyId":"wNI9atQGlz+VWfO6LRygH4QUfY/8W4RFwiT5i5WRgB0="}, "kindVersion":{"kind":"hashedrekord", "version":"0.0.1"}, "integratedTime":"1792423926", "inclusionPromise":{"signedEntryTimestamp":"MEQCIBtXxI5gheG1qsE9cX71H3Q32jllTixEh8L+StwBj7L4AiAWragGxKJY4X8oPcTmxNbDfqff4kvWV+Z46Ev2qwXU1Q=="}, "inclusionProof":{"logIndex":"3200800602", "rootHash":"IiO0LKB3IFknrU6UpIr7ODvebLbNOEdj9CWSMoa4/9E=", "treeSize":"3200800672", "hashes":["NwPGfsCmi2VTVGnlaOQOVqU/qj/C7tvybvUrdCzLISM=", "tbKj5qrhpk4NMeMqwIkJP3oB1E3L0Xf8hfJRP7ErodU=", "BUHt9OGVfND4xwHLLlL9k6QYLFHEU4A318uAADEXOcc=", "P3z3QkOScUwyzyJARHjYbf9YxpEqyXAGCK2TX+MhDBI=", "9Vxsyt/Fgxo23kUkhRTK0i8hIh8DcrlyrB2iqU+06wo=", "LP7vlIJyS9EitsBtRs7FH/YJHROdcYFx75BcthGOmcg=", "5v00gnLtujY6FeREDLWauBC5tFyHTU3Z/6deTmWPTIE=", "Z4UjocBt9yeNh1uonx3cqDqIKCXzAdDQwqZDQoCiW9k=", "vXpDygpx9rhN9LXfidYB+BPwgAn4PMuXykP8b3bKj7c=", "Le0PNZGukT+JaApRQKno8TysO4tvutud1B2LUyPWOd8=", "Y/kTXkjjF16KI/qmcw9vLu8ZBoFtGZRa17fGXCGRAp0=", "11lr1PAKx8EF0MQ8eyTG2Okak7VNiBdR4LabBugomSw=", "uGEIH4dmv632MUSdl5AVDq3lFz9EPX5Y4BnnOA6WyoM=", "D37E7MwfvuENRJ4BUsqOTbzKKFjz+Dkh4kKr7xMQyU8=", "7WFpoaiQCzLwM3o2YI5Ngo8J7JD0OfyCCQtmXhWxVes=", "nyQV/+cWQUG5Hb2PKjQQNkQIJmz2qzAZKBQlkHGIknM=", "G+nk7mgE5h6dUmv2UY7NDJrY3WvrZY31raaO1AETt+o=", "qlTWNDzERnTBQVcHHXXVPjiUP0SsikGxfcxVlgqVFDo=", "gg9GNApNke3+zNhl4tsWtmIycCFrOMCB3lys6XuSQNQ=", "NlOclldmKPA1tMKIskqHAwuQDYHp31YwpK+OOMrMccc=", "qxzHanAzz57SDdmJe0B7bJK72NTIBbwEMGBBKvDOROw=", "xH/DCseLHr9eKoYT8qsORZK7zVdEGYWHuVtsVrD95wY="], "checkpoint":{"envelope":"rekor.sigstore.dev - 1193050959916656506\n3200800672\nIiO0LKB3IFknrU6UpIr7ODvebLbNOEdj9CWSMoa4/9E=\n\n— rekor.sigstore.dev wNI9ajBFAiEAv9Fkefugg9FPKSdsTW8afhDdkJAnPgOxL+6PLrQRPOcCIAXLnKdyxggYWvD9Kx8sajLcEkVh3jjRmVGR+ZbYe5OG\n"}}, "canonicalizedBody":"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"}], "timestampVerificationData":{"rfc3161Timestamps":[{"signedTimestamp":"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"}]}}, "messageSignature":{"messageDigest":{"algorithm":"SHA2_256", "digest":"ZsBEkauQlBeEiplTk/Ys0huE9daDp6g5GWUnjbqAiTk="}, "signature":"MEUCIHV3Wi5iyCCHAlX79Jg2xYMEIFDwo3KrjgXEZwP1sVaCAiEAymK2gr8ZEE9hcJiCUVb1UilLzngVyYN+NAGCac78zjc="}}
|
build/torch-stable-abi29-cu128-x86_64-linux/symbols.json
ADDED
|
@@ -0,0 +1,745 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"schema_version": 1,
|
| 3 |
+
"module": "_flash_attn3_cuda_fe2bb66",
|
| 4 |
+
"functions": [
|
| 5 |
+
{
|
| 6 |
+
"name": "add_op_namespace_prefix",
|
| 7 |
+
"qualname": "add_op_namespace_prefix",
|
| 8 |
+
"module": "_flash_attn3_cuda_fe2bb66._ops",
|
| 9 |
+
"kind": "function",
|
| 10 |
+
"doc": "Prefix op by namespace.",
|
| 11 |
+
"signature": {
|
| 12 |
+
"parameters": [
|
| 13 |
+
{
|
| 14 |
+
"name": "op_name",
|
| 15 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 16 |
+
"default": null,
|
| 17 |
+
"annotation": "str"
|
| 18 |
+
}
|
| 19 |
+
],
|
| 20 |
+
"return_annotation": null
|
| 21 |
+
}
|
| 22 |
+
},
|
| 23 |
+
{
|
| 24 |
+
"name": "flash_attn_combine",
|
| 25 |
+
"qualname": "flash_attn_combine",
|
| 26 |
+
"module": "_flash_attn3_cuda_fe2bb66.flash_attn_interface",
|
| 27 |
+
"kind": "function",
|
| 28 |
+
"doc": null,
|
| 29 |
+
"signature": {
|
| 30 |
+
"parameters": [
|
| 31 |
+
{
|
| 32 |
+
"name": "out_partial",
|
| 33 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 34 |
+
"default": null,
|
| 35 |
+
"annotation": null
|
| 36 |
+
},
|
| 37 |
+
{
|
| 38 |
+
"name": "lse_partial",
|
| 39 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 40 |
+
"default": null,
|
| 41 |
+
"annotation": null
|
| 42 |
+
},
|
| 43 |
+
{
|
| 44 |
+
"name": "out",
|
| 45 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 46 |
+
"default": "None",
|
| 47 |
+
"annotation": null
|
| 48 |
+
},
|
| 49 |
+
{
|
| 50 |
+
"name": "out_dtype",
|
| 51 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 52 |
+
"default": "None",
|
| 53 |
+
"annotation": null
|
| 54 |
+
}
|
| 55 |
+
],
|
| 56 |
+
"return_annotation": null
|
| 57 |
+
}
|
| 58 |
+
},
|
| 59 |
+
{
|
| 60 |
+
"name": "flash_attn_func",
|
| 61 |
+
"qualname": "flash_attn_func",
|
| 62 |
+
"module": "_flash_attn3_cuda_fe2bb66.flash_attn_interface",
|
| 63 |
+
"kind": "function",
|
| 64 |
+
"doc": "dropout_p should be set to 0.0 during evaluation\nSupports multi-query and grouped-query attention (MQA/GQA) by passing in KV with fewer heads\nthan Q. Note that the number of heads in Q must be divisible by the number of heads in KV.\nFor example, if Q has 6 heads and K, V have 2 heads, head 0, 1, 2 of Q will attention to head\n0 of K, V, and head 3, 4, 5 of Q will attention to head 1 of K, V.\n\nIf causal=True, the causal mask is aligned to the bottom right corner of the attention matrix.\nFor example, if seqlen_q = 2 and seqlen_k = 5, the causal mask (1 = keep, 0 = masked out) is:\n 1 1 1 1 0\n 1 1 1 1 1\nIf seqlen_q = 5 and seqlen_k = 2, the causal mask is:\n 0 0\n 0 0\n 0 0\n 1 0\n 1 1\nIf the row of the mask is all zero, the output will be zero.\n\nIf window_size != (-1, -1), implements sliding window local attention. Query at position i\nwill only attend to keys between\n[i + seqlen_k - seqlen_q - window_size[0], i + seqlen_k - seqlen_q + window_size[1]] inclusive.\n\nArguments:\n q: (batch_size, seqlen, nheads, headdim)\n k: (batch_size, seqlen, nheads_k, headdim)\n v: (batch_size, seqlen, nheads_k, headdim)\n dropout_p: float. Dropout probability.\n softmax_scale: float. The scaling of QK^T before applying softmax.\n Default to 1 / sqrt(headdim).\n causal: bool. Whether to apply causal attention mask (e.g., for auto-regressive modeling).\n window_size: (left, right). If not (-1, -1), implements sliding window local attention.\n alibi_slopes: (nheads,) or (batch_size, nheads), fp32. A bias of\n (-alibi_slope * |i + seqlen_k - seqlen_q - j|)\n is added to the attention score of query i and key j.\n deterministic: bool. Whether to use the deterministic implementation of the backward pass,\n which is slightly slower and uses more memory. The forward pass is always deterministic.\n return_attn_probs: bool. Whether to return the attention probabilities. This option is for\n testing only. The returned probabilities are not guaranteed to be correct\n (they might not have the right scaling).\nReturn:\n out: (batch_size, seqlen, nheads, headdim).\n softmax_lse [optional, if return_attn_probs=True]: (batch_size, nheads, seqlen). The\n logsumexp of each row of the matrix QK^T * scaling (e.g., log of the softmax\n normalization factor).",
|
| 65 |
+
"signature": {
|
| 66 |
+
"parameters": [
|
| 67 |
+
{
|
| 68 |
+
"name": "q",
|
| 69 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 70 |
+
"default": null,
|
| 71 |
+
"annotation": null
|
| 72 |
+
},
|
| 73 |
+
{
|
| 74 |
+
"name": "k",
|
| 75 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 76 |
+
"default": null,
|
| 77 |
+
"annotation": null
|
| 78 |
+
},
|
| 79 |
+
{
|
| 80 |
+
"name": "v",
|
| 81 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 82 |
+
"default": null,
|
| 83 |
+
"annotation": null
|
| 84 |
+
},
|
| 85 |
+
{
|
| 86 |
+
"name": "softmax_scale",
|
| 87 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 88 |
+
"default": "None",
|
| 89 |
+
"annotation": null
|
| 90 |
+
},
|
| 91 |
+
{
|
| 92 |
+
"name": "causal",
|
| 93 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 94 |
+
"default": "False",
|
| 95 |
+
"annotation": null
|
| 96 |
+
},
|
| 97 |
+
{
|
| 98 |
+
"name": "qv",
|
| 99 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 100 |
+
"default": "None",
|
| 101 |
+
"annotation": null
|
| 102 |
+
},
|
| 103 |
+
{
|
| 104 |
+
"name": "q_descale",
|
| 105 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 106 |
+
"default": "None",
|
| 107 |
+
"annotation": null
|
| 108 |
+
},
|
| 109 |
+
{
|
| 110 |
+
"name": "k_descale",
|
| 111 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 112 |
+
"default": "None",
|
| 113 |
+
"annotation": null
|
| 114 |
+
},
|
| 115 |
+
{
|
| 116 |
+
"name": "v_descale",
|
| 117 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 118 |
+
"default": "None",
|
| 119 |
+
"annotation": null
|
| 120 |
+
},
|
| 121 |
+
{
|
| 122 |
+
"name": "window_size",
|
| 123 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 124 |
+
"default": "(-1, -1)",
|
| 125 |
+
"annotation": null
|
| 126 |
+
},
|
| 127 |
+
{
|
| 128 |
+
"name": "attention_chunk",
|
| 129 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 130 |
+
"default": "0",
|
| 131 |
+
"annotation": null
|
| 132 |
+
},
|
| 133 |
+
{
|
| 134 |
+
"name": "softcap",
|
| 135 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 136 |
+
"default": "0.0",
|
| 137 |
+
"annotation": null
|
| 138 |
+
},
|
| 139 |
+
{
|
| 140 |
+
"name": "num_splits",
|
| 141 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 142 |
+
"default": "1",
|
| 143 |
+
"annotation": null
|
| 144 |
+
},
|
| 145 |
+
{
|
| 146 |
+
"name": "pack_gqa",
|
| 147 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 148 |
+
"default": "None",
|
| 149 |
+
"annotation": null
|
| 150 |
+
},
|
| 151 |
+
{
|
| 152 |
+
"name": "deterministic",
|
| 153 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 154 |
+
"default": "False",
|
| 155 |
+
"annotation": null
|
| 156 |
+
},
|
| 157 |
+
{
|
| 158 |
+
"name": "sm_margin",
|
| 159 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 160 |
+
"default": "0",
|
| 161 |
+
"annotation": null
|
| 162 |
+
},
|
| 163 |
+
{
|
| 164 |
+
"name": "return_attn_probs",
|
| 165 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 166 |
+
"default": "False",
|
| 167 |
+
"annotation": null
|
| 168 |
+
}
|
| 169 |
+
],
|
| 170 |
+
"return_annotation": null
|
| 171 |
+
}
|
| 172 |
+
},
|
| 173 |
+
{
|
| 174 |
+
"name": "flash_attn_qkvpacked_func",
|
| 175 |
+
"qualname": "flash_attn_qkvpacked_func",
|
| 176 |
+
"module": "_flash_attn3_cuda_fe2bb66.flash_attn_interface",
|
| 177 |
+
"kind": "function",
|
| 178 |
+
"doc": "dropout_p should be set to 0.0 during evaluation\nIf Q, K, V are already stacked into 1 tensor, this function will be faster than\ncalling flash_attn_func on Q, K, V since the backward pass avoids explicit concatenation\nof the gradients of Q, K, V.\nFor multi-query and grouped-query attention (MQA/GQA), please see\nflash_attn_kvpacked_func and flash_attn_func.\n\nIf window_size != (-1, -1), implements sliding window local attention. Query at position i\nwill only attend to keys between [i - window_size[0], i + window_size[1]] inclusive.\n\nArguments:\n qkv: (batch_size, seqlen, 3, nheads, headdim)\n dropout_p: float. Dropout probability.\n softmax_scale: float. The scaling of QK^T before applying softmax.\n Default to 1 / sqrt(headdim).\n causal: bool. Whether to apply causal attention mask (e.g., for auto-regressive modeling).\n window_size: (left, right). If not (-1, -1), implements sliding window local attention.\n softcap: float. Anything > 0 activates softcapping attention.\n alibi_slopes: (nheads,) or (batch_size, nheads), fp32. A bias of (-alibi_slope * |i - j|) is added to\n the attention score of query i and key j.\n deterministic: bool. Whether to use the deterministic implementation of the backward pass,\n which is slightly slower and uses more memory. The forward pass is always deterministic.\n return_attn_probs: bool. Whether to return the attention probabilities. This option is for\n testing only. The returned probabilities are not guaranteed to be correct\n (they might not have the right scaling).\nReturn:\n out: (batch_size, seqlen, nheads, headdim).\n softmax_lse [optional, if return_attn_probs=True]: (batch_size, nheads, seqlen). The\n logsumexp of each row of the matrix QK^T * scaling (e.g., log of the softmax\n normalization factor).\n S_dmask [optional, if return_attn_probs=True]: (batch_size, nheads, seqlen, seqlen).\n The output of softmax (possibly with different scaling). It also encodes the dropout\n pattern (negative means that location was dropped, nonnegative means it was kept).",
|
| 179 |
+
"signature": {
|
| 180 |
+
"parameters": [
|
| 181 |
+
{
|
| 182 |
+
"name": "qkv",
|
| 183 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 184 |
+
"default": null,
|
| 185 |
+
"annotation": null
|
| 186 |
+
},
|
| 187 |
+
{
|
| 188 |
+
"name": "softmax_scale",
|
| 189 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 190 |
+
"default": "None",
|
| 191 |
+
"annotation": null
|
| 192 |
+
},
|
| 193 |
+
{
|
| 194 |
+
"name": "causal",
|
| 195 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 196 |
+
"default": "False",
|
| 197 |
+
"annotation": null
|
| 198 |
+
},
|
| 199 |
+
{
|
| 200 |
+
"name": "q_descale",
|
| 201 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 202 |
+
"default": "None",
|
| 203 |
+
"annotation": null
|
| 204 |
+
},
|
| 205 |
+
{
|
| 206 |
+
"name": "k_descale",
|
| 207 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 208 |
+
"default": "None",
|
| 209 |
+
"annotation": null
|
| 210 |
+
},
|
| 211 |
+
{
|
| 212 |
+
"name": "v_descale",
|
| 213 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 214 |
+
"default": "None",
|
| 215 |
+
"annotation": null
|
| 216 |
+
},
|
| 217 |
+
{
|
| 218 |
+
"name": "window_size",
|
| 219 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 220 |
+
"default": "(-1, -1)",
|
| 221 |
+
"annotation": null
|
| 222 |
+
},
|
| 223 |
+
{
|
| 224 |
+
"name": "attention_chunk",
|
| 225 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 226 |
+
"default": "0",
|
| 227 |
+
"annotation": null
|
| 228 |
+
},
|
| 229 |
+
{
|
| 230 |
+
"name": "softcap",
|
| 231 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 232 |
+
"default": "0.0",
|
| 233 |
+
"annotation": null
|
| 234 |
+
},
|
| 235 |
+
{
|
| 236 |
+
"name": "deterministic",
|
| 237 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 238 |
+
"default": "False",
|
| 239 |
+
"annotation": null
|
| 240 |
+
},
|
| 241 |
+
{
|
| 242 |
+
"name": "num_heads_q",
|
| 243 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 244 |
+
"default": "None",
|
| 245 |
+
"annotation": null
|
| 246 |
+
},
|
| 247 |
+
{
|
| 248 |
+
"name": "sm_margin",
|
| 249 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 250 |
+
"default": "0",
|
| 251 |
+
"annotation": null
|
| 252 |
+
},
|
| 253 |
+
{
|
| 254 |
+
"name": "return_attn_probs",
|
| 255 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 256 |
+
"default": "False",
|
| 257 |
+
"annotation": null
|
| 258 |
+
}
|
| 259 |
+
],
|
| 260 |
+
"return_annotation": null
|
| 261 |
+
}
|
| 262 |
+
},
|
| 263 |
+
{
|
| 264 |
+
"name": "flash_attn_varlen_func",
|
| 265 |
+
"qualname": "flash_attn_varlen_func",
|
| 266 |
+
"module": "_flash_attn3_cuda_fe2bb66.flash_attn_interface",
|
| 267 |
+
"kind": "function",
|
| 268 |
+
"doc": null,
|
| 269 |
+
"signature": {
|
| 270 |
+
"parameters": [
|
| 271 |
+
{
|
| 272 |
+
"name": "q",
|
| 273 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 274 |
+
"default": null,
|
| 275 |
+
"annotation": null
|
| 276 |
+
},
|
| 277 |
+
{
|
| 278 |
+
"name": "k",
|
| 279 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 280 |
+
"default": null,
|
| 281 |
+
"annotation": null
|
| 282 |
+
},
|
| 283 |
+
{
|
| 284 |
+
"name": "v",
|
| 285 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 286 |
+
"default": null,
|
| 287 |
+
"annotation": null
|
| 288 |
+
},
|
| 289 |
+
{
|
| 290 |
+
"name": "cu_seqlens_q",
|
| 291 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 292 |
+
"default": null,
|
| 293 |
+
"annotation": null
|
| 294 |
+
},
|
| 295 |
+
{
|
| 296 |
+
"name": "cu_seqlens_k",
|
| 297 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 298 |
+
"default": null,
|
| 299 |
+
"annotation": null
|
| 300 |
+
},
|
| 301 |
+
{
|
| 302 |
+
"name": "max_seqlen_q",
|
| 303 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 304 |
+
"default": null,
|
| 305 |
+
"annotation": null
|
| 306 |
+
},
|
| 307 |
+
{
|
| 308 |
+
"name": "max_seqlen_k",
|
| 309 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 310 |
+
"default": null,
|
| 311 |
+
"annotation": null
|
| 312 |
+
},
|
| 313 |
+
{
|
| 314 |
+
"name": "seqused_q",
|
| 315 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 316 |
+
"default": "None",
|
| 317 |
+
"annotation": null
|
| 318 |
+
},
|
| 319 |
+
{
|
| 320 |
+
"name": "seqused_k",
|
| 321 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 322 |
+
"default": "None",
|
| 323 |
+
"annotation": null
|
| 324 |
+
},
|
| 325 |
+
{
|
| 326 |
+
"name": "softmax_scale",
|
| 327 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 328 |
+
"default": "None",
|
| 329 |
+
"annotation": null
|
| 330 |
+
},
|
| 331 |
+
{
|
| 332 |
+
"name": "causal",
|
| 333 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 334 |
+
"default": "False",
|
| 335 |
+
"annotation": null
|
| 336 |
+
},
|
| 337 |
+
{
|
| 338 |
+
"name": "qv",
|
| 339 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 340 |
+
"default": "None",
|
| 341 |
+
"annotation": null
|
| 342 |
+
},
|
| 343 |
+
{
|
| 344 |
+
"name": "q_descale",
|
| 345 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 346 |
+
"default": "None",
|
| 347 |
+
"annotation": null
|
| 348 |
+
},
|
| 349 |
+
{
|
| 350 |
+
"name": "k_descale",
|
| 351 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 352 |
+
"default": "None",
|
| 353 |
+
"annotation": null
|
| 354 |
+
},
|
| 355 |
+
{
|
| 356 |
+
"name": "v_descale",
|
| 357 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 358 |
+
"default": "None",
|
| 359 |
+
"annotation": null
|
| 360 |
+
},
|
| 361 |
+
{
|
| 362 |
+
"name": "window_size",
|
| 363 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 364 |
+
"default": "(-1, -1)",
|
| 365 |
+
"annotation": null
|
| 366 |
+
},
|
| 367 |
+
{
|
| 368 |
+
"name": "attention_chunk",
|
| 369 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 370 |
+
"default": "0",
|
| 371 |
+
"annotation": null
|
| 372 |
+
},
|
| 373 |
+
{
|
| 374 |
+
"name": "softcap",
|
| 375 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 376 |
+
"default": "0.0",
|
| 377 |
+
"annotation": null
|
| 378 |
+
},
|
| 379 |
+
{
|
| 380 |
+
"name": "num_splits",
|
| 381 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 382 |
+
"default": "1",
|
| 383 |
+
"annotation": null
|
| 384 |
+
},
|
| 385 |
+
{
|
| 386 |
+
"name": "pack_gqa",
|
| 387 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 388 |
+
"default": "None",
|
| 389 |
+
"annotation": null
|
| 390 |
+
},
|
| 391 |
+
{
|
| 392 |
+
"name": "deterministic",
|
| 393 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 394 |
+
"default": "False",
|
| 395 |
+
"annotation": null
|
| 396 |
+
},
|
| 397 |
+
{
|
| 398 |
+
"name": "sm_margin",
|
| 399 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 400 |
+
"default": "0",
|
| 401 |
+
"annotation": null
|
| 402 |
+
},
|
| 403 |
+
{
|
| 404 |
+
"name": "return_attn_probs",
|
| 405 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 406 |
+
"default": "False",
|
| 407 |
+
"annotation": null
|
| 408 |
+
}
|
| 409 |
+
],
|
| 410 |
+
"return_annotation": null
|
| 411 |
+
}
|
| 412 |
+
},
|
| 413 |
+
{
|
| 414 |
+
"name": "flash_attn_with_kvcache",
|
| 415 |
+
"qualname": "flash_attn_with_kvcache",
|
| 416 |
+
"module": "_flash_attn3_cuda_fe2bb66.flash_attn_interface",
|
| 417 |
+
"kind": "function",
|
| 418 |
+
"doc": "If k and v are not None, k_cache and v_cache will be updated *inplace* with the new values from\nk and v. This is useful for incremental decoding: you can pass in the cached keys/values from\nthe previous step, and update them with the new keys/values from the current step, and do\nattention with the updated cache, all in 1 kernel.\n\nIf you pass in k / v, you must make sure that the cache is large enough to hold the new values.\nFor example, the KV cache could be pre-allocated with the max sequence length, and you can use\ncache_seqlens to keep track of the current sequence lengths of each sequence in the batch.\n\nAlso apply rotary embedding if rotary_cos and rotary_sin are passed in. The key @k will be\nrotated by rotary_cos and rotary_sin at indices cache_seqlens, cache_seqlens + 1, etc.\nIf causal or local (i.e., window_size != (-1, -1)), the query @q will be rotated by rotary_cos\nand rotary_sin at indices cache_seqlens, cache_seqlens + 1, etc.\nIf not causal and not local, the query @q will be rotated by rotary_cos and rotary_sin at\nindices cache_seqlens only (i.e. we consider all tokens in @q to be at position cache_seqlens).\n\nSee tests/test_flash_attn.py::test_flash_attn_kvcache for examples of how to use this function.\n\nSupports multi-query and grouped-query attention (MQA/GQA) by passing in KV with fewer heads\nthan Q. Note that the number of heads in Q must be divisible by the number of heads in KV.\nFor example, if Q has 6 heads and K, V have 2 heads, head 0, 1, 2 of Q will attention to head\n0 of K, V, and head 3, 4, 5 of Q will attention to head 1 of K, V.\n\nIf causal=True, the causal mask is aligned to the bottom right corner of the attention matrix.\nFor example, if seqlen_q = 2 and seqlen_k = 5, the causal mask (1 = keep, 0 = masked out) is:\n 1 1 1 1 0\n 1 1 1 1 1\nIf seqlen_q = 5 and seqlen_k = 2, the causal mask is:\n 0 0\n 0 0\n 0 0\n 1 0\n 1 1\nIf the row of the mask is all zero, the output will be zero.\n\nIf window_size != (-1, -1), implements sliding window local attention. Query at position i\nwill only attend to keys between\n[i + seqlen_k - seqlen_q - window_size[0], i + seqlen_k - seqlen_q + window_size[1]] inclusive.\n\nNote: Does not support backward pass.\n\nArguments:\n q: (batch_size, seqlen, nheads, headdim)\n k_cache: (batch_size_cache, seqlen_cache, nheads_k, headdim) if there's no page_table,\n or (num_blocks, page_block_size, nheads_k, headdim) if there's a page_table (i.e. paged KV cache)\n page_block_size can be arbitrary (e.g, 1, 2, 3, 64, etc.).\n v_cache: (batch_size_cache, seqlen_cache, nheads_k, headdim_v) if there's no page_table,\n or (num_blocks, page_block_size, nheads_k, headdim_v) if there's a page_table (i.e. paged KV cache)\n k [optional]: (batch_size, seqlen_new, nheads_k, headdim). If not None, we concatenate\n k with k_cache, starting at the indices specified by cache_seqlens.\n v [optional]: (batch_size, seqlen_new, nheads_k, headdim_v). Similar to k.\n qv [optional]: (batch_size, seqlen, nheads, headdim_v)\n rotary_cos [optional]: (seqlen_ro, rotary_dim / 2). If not None, we apply rotary embedding\n to k and q. Only applicable if k and v are passed in. rotary_dim must be divisible by 16.\n rotary_sin [optional]: (seqlen_ro, rotary_dim / 2). Similar to rotary_cos.\n cache_seqlens: int, or (batch_size,), dtype torch.int32. The sequence lengths of the\n KV cache.\n cache_batch_idx: (batch_size,), dtype torch.int32. The indices used to index into the KV cache.\n If None, we assume that the batch indices are [0, 1, 2, ..., batch_size - 1].\n If the indices are not distinct, and k and v are provided, the values updated in the cache\n might come from any of the duplicate indices.\n cache_leftpad: (batch_size,), dtype torch.int32. The index that the KV cache starts. If None, assume 0.\n page_table [optional]: (batch_size, max_num_blocks_per_seq), dtype torch.int32.\n softmax_scale: float. The scaling of QK^T before applying softmax.\n Default to 1 / sqrt(headdim).\n causal: bool. Whether to apply causal attention mask (e.g., for auto-regressive modeling).\n window_size: (left, right). If not (-1, -1), implements sliding window local attention.\n softcap: float. Anything > 0 activates softcapping attention.\n rotary_interleaved: bool. Only applicable if rotary_cos and rotary_sin are passed in.\n If True, rotary embedding will combine dimensions 0 & 1, 2 & 3, etc. If False,\n rotary embedding will combine dimensions 0 & rotary_dim / 2, 1 & rotary_dim / 2 + 1\n (i.e. GPT-NeoX style).\n num_splits: int. If > 1, split the key/value into this many chunks along the sequence.\n If num_splits == 1, we don't split the key/value. If num_splits == 0, we use a heuristic\n to automatically determine the number of splits.\n Don't change this unless you know what you are doing.\n return_softmax_lse: bool. Whether to return the logsumexp of the attention scores.\n\nReturn:\n out: (batch_size, seqlen, nheads, headdim).\n softmax_lse [optional, if return_softmax_lse=True]: (batch_size, nheads, seqlen). The\n logsumexp of each row of the matrix QK^T * scaling (e.g., log of the softmax\n normalization factor).",
|
| 419 |
+
"signature": {
|
| 420 |
+
"parameters": [
|
| 421 |
+
{
|
| 422 |
+
"name": "q",
|
| 423 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 424 |
+
"default": null,
|
| 425 |
+
"annotation": null
|
| 426 |
+
},
|
| 427 |
+
{
|
| 428 |
+
"name": "k_cache",
|
| 429 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 430 |
+
"default": null,
|
| 431 |
+
"annotation": null
|
| 432 |
+
},
|
| 433 |
+
{
|
| 434 |
+
"name": "v_cache",
|
| 435 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 436 |
+
"default": null,
|
| 437 |
+
"annotation": null
|
| 438 |
+
},
|
| 439 |
+
{
|
| 440 |
+
"name": "k",
|
| 441 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 442 |
+
"default": "None",
|
| 443 |
+
"annotation": null
|
| 444 |
+
},
|
| 445 |
+
{
|
| 446 |
+
"name": "v",
|
| 447 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 448 |
+
"default": "None",
|
| 449 |
+
"annotation": null
|
| 450 |
+
},
|
| 451 |
+
{
|
| 452 |
+
"name": "qv",
|
| 453 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 454 |
+
"default": "None",
|
| 455 |
+
"annotation": null
|
| 456 |
+
},
|
| 457 |
+
{
|
| 458 |
+
"name": "rotary_cos",
|
| 459 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 460 |
+
"default": "None",
|
| 461 |
+
"annotation": null
|
| 462 |
+
},
|
| 463 |
+
{
|
| 464 |
+
"name": "rotary_sin",
|
| 465 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 466 |
+
"default": "None",
|
| 467 |
+
"annotation": null
|
| 468 |
+
},
|
| 469 |
+
{
|
| 470 |
+
"name": "cache_seqlens",
|
| 471 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 472 |
+
"default": "None",
|
| 473 |
+
"annotation": "int | torch.Tensor | None"
|
| 474 |
+
},
|
| 475 |
+
{
|
| 476 |
+
"name": "cache_batch_idx",
|
| 477 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 478 |
+
"default": "None",
|
| 479 |
+
"annotation": "torch.Tensor | None"
|
| 480 |
+
},
|
| 481 |
+
{
|
| 482 |
+
"name": "cache_leftpad",
|
| 483 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 484 |
+
"default": "None",
|
| 485 |
+
"annotation": "torch.Tensor | None"
|
| 486 |
+
},
|
| 487 |
+
{
|
| 488 |
+
"name": "page_table",
|
| 489 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 490 |
+
"default": "None",
|
| 491 |
+
"annotation": "torch.Tensor | None"
|
| 492 |
+
},
|
| 493 |
+
{
|
| 494 |
+
"name": "cu_seqlens_q",
|
| 495 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 496 |
+
"default": "None",
|
| 497 |
+
"annotation": "torch.Tensor | None"
|
| 498 |
+
},
|
| 499 |
+
{
|
| 500 |
+
"name": "cu_seqlens_k_new",
|
| 501 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 502 |
+
"default": "None",
|
| 503 |
+
"annotation": "torch.Tensor | None"
|
| 504 |
+
},
|
| 505 |
+
{
|
| 506 |
+
"name": "max_seqlen_q",
|
| 507 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 508 |
+
"default": "None",
|
| 509 |
+
"annotation": "int | None"
|
| 510 |
+
},
|
| 511 |
+
{
|
| 512 |
+
"name": "rotary_seqlens",
|
| 513 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 514 |
+
"default": "None",
|
| 515 |
+
"annotation": "torch.Tensor | None"
|
| 516 |
+
},
|
| 517 |
+
{
|
| 518 |
+
"name": "q_descale",
|
| 519 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 520 |
+
"default": "None",
|
| 521 |
+
"annotation": "torch.Tensor | None"
|
| 522 |
+
},
|
| 523 |
+
{
|
| 524 |
+
"name": "k_descale",
|
| 525 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 526 |
+
"default": "None",
|
| 527 |
+
"annotation": "torch.Tensor | None"
|
| 528 |
+
},
|
| 529 |
+
{
|
| 530 |
+
"name": "v_descale",
|
| 531 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 532 |
+
"default": "None",
|
| 533 |
+
"annotation": "torch.Tensor | None"
|
| 534 |
+
},
|
| 535 |
+
{
|
| 536 |
+
"name": "softmax_scale",
|
| 537 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 538 |
+
"default": "None",
|
| 539 |
+
"annotation": null
|
| 540 |
+
},
|
| 541 |
+
{
|
| 542 |
+
"name": "causal",
|
| 543 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 544 |
+
"default": "False",
|
| 545 |
+
"annotation": null
|
| 546 |
+
},
|
| 547 |
+
{
|
| 548 |
+
"name": "window_size",
|
| 549 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 550 |
+
"default": "(-1, -1)",
|
| 551 |
+
"annotation": null
|
| 552 |
+
},
|
| 553 |
+
{
|
| 554 |
+
"name": "attention_chunk",
|
| 555 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 556 |
+
"default": "0",
|
| 557 |
+
"annotation": null
|
| 558 |
+
},
|
| 559 |
+
{
|
| 560 |
+
"name": "softcap",
|
| 561 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 562 |
+
"default": "0.0",
|
| 563 |
+
"annotation": null
|
| 564 |
+
},
|
| 565 |
+
{
|
| 566 |
+
"name": "rotary_interleaved",
|
| 567 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 568 |
+
"default": "True",
|
| 569 |
+
"annotation": null
|
| 570 |
+
},
|
| 571 |
+
{
|
| 572 |
+
"name": "scheduler_metadata",
|
| 573 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 574 |
+
"default": "None",
|
| 575 |
+
"annotation": null
|
| 576 |
+
},
|
| 577 |
+
{
|
| 578 |
+
"name": "num_splits",
|
| 579 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 580 |
+
"default": "0",
|
| 581 |
+
"annotation": null
|
| 582 |
+
},
|
| 583 |
+
{
|
| 584 |
+
"name": "pack_gqa",
|
| 585 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 586 |
+
"default": "None",
|
| 587 |
+
"annotation": null
|
| 588 |
+
},
|
| 589 |
+
{
|
| 590 |
+
"name": "sm_margin",
|
| 591 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 592 |
+
"default": "0",
|
| 593 |
+
"annotation": null
|
| 594 |
+
},
|
| 595 |
+
{
|
| 596 |
+
"name": "return_softmax_lse",
|
| 597 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 598 |
+
"default": "False",
|
| 599 |
+
"annotation": null
|
| 600 |
+
}
|
| 601 |
+
],
|
| 602 |
+
"return_annotation": null
|
| 603 |
+
}
|
| 604 |
+
},
|
| 605 |
+
{
|
| 606 |
+
"name": "get_scheduler_metadata",
|
| 607 |
+
"qualname": "get_scheduler_metadata",
|
| 608 |
+
"module": "_flash_attn3_cuda_fe2bb66.flash_attn_interface",
|
| 609 |
+
"kind": "function",
|
| 610 |
+
"doc": null,
|
| 611 |
+
"signature": {
|
| 612 |
+
"parameters": [
|
| 613 |
+
{
|
| 614 |
+
"name": "batch_size",
|
| 615 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 616 |
+
"default": null,
|
| 617 |
+
"annotation": null
|
| 618 |
+
},
|
| 619 |
+
{
|
| 620 |
+
"name": "max_seqlen_q",
|
| 621 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 622 |
+
"default": null,
|
| 623 |
+
"annotation": null
|
| 624 |
+
},
|
| 625 |
+
{
|
| 626 |
+
"name": "max_seqlen_k",
|
| 627 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 628 |
+
"default": null,
|
| 629 |
+
"annotation": null
|
| 630 |
+
},
|
| 631 |
+
{
|
| 632 |
+
"name": "num_heads_q",
|
| 633 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 634 |
+
"default": null,
|
| 635 |
+
"annotation": null
|
| 636 |
+
},
|
| 637 |
+
{
|
| 638 |
+
"name": "num_heads_kv",
|
| 639 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 640 |
+
"default": null,
|
| 641 |
+
"annotation": null
|
| 642 |
+
},
|
| 643 |
+
{
|
| 644 |
+
"name": "headdim",
|
| 645 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 646 |
+
"default": null,
|
| 647 |
+
"annotation": null
|
| 648 |
+
},
|
| 649 |
+
{
|
| 650 |
+
"name": "cache_seqlens",
|
| 651 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 652 |
+
"default": null,
|
| 653 |
+
"annotation": "torch.Tensor"
|
| 654 |
+
},
|
| 655 |
+
{
|
| 656 |
+
"name": "qkv_dtype",
|
| 657 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 658 |
+
"default": "torch.bfloat16",
|
| 659 |
+
"annotation": null
|
| 660 |
+
},
|
| 661 |
+
{
|
| 662 |
+
"name": "headdim_v",
|
| 663 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 664 |
+
"default": "None",
|
| 665 |
+
"annotation": null
|
| 666 |
+
},
|
| 667 |
+
{
|
| 668 |
+
"name": "cu_seqlens_q",
|
| 669 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 670 |
+
"default": "None",
|
| 671 |
+
"annotation": "torch.Tensor | None"
|
| 672 |
+
},
|
| 673 |
+
{
|
| 674 |
+
"name": "cu_seqlens_k_new",
|
| 675 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 676 |
+
"default": "None",
|
| 677 |
+
"annotation": "torch.Tensor | None"
|
| 678 |
+
},
|
| 679 |
+
{
|
| 680 |
+
"name": "cache_leftpad",
|
| 681 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 682 |
+
"default": "None",
|
| 683 |
+
"annotation": "torch.Tensor | None"
|
| 684 |
+
},
|
| 685 |
+
{
|
| 686 |
+
"name": "page_size",
|
| 687 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 688 |
+
"default": "None",
|
| 689 |
+
"annotation": "int | None"
|
| 690 |
+
},
|
| 691 |
+
{
|
| 692 |
+
"name": "max_seqlen_k_new",
|
| 693 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 694 |
+
"default": "0",
|
| 695 |
+
"annotation": null
|
| 696 |
+
},
|
| 697 |
+
{
|
| 698 |
+
"name": "causal",
|
| 699 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 700 |
+
"default": "False",
|
| 701 |
+
"annotation": null
|
| 702 |
+
},
|
| 703 |
+
{
|
| 704 |
+
"name": "window_size",
|
| 705 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 706 |
+
"default": "(-1, -1)",
|
| 707 |
+
"annotation": null
|
| 708 |
+
},
|
| 709 |
+
{
|
| 710 |
+
"name": "attention_chunk",
|
| 711 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 712 |
+
"default": "0",
|
| 713 |
+
"annotation": null
|
| 714 |
+
},
|
| 715 |
+
{
|
| 716 |
+
"name": "has_softcap",
|
| 717 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 718 |
+
"default": "False",
|
| 719 |
+
"annotation": null
|
| 720 |
+
},
|
| 721 |
+
{
|
| 722 |
+
"name": "num_splits",
|
| 723 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 724 |
+
"default": "0",
|
| 725 |
+
"annotation": null
|
| 726 |
+
},
|
| 727 |
+
{
|
| 728 |
+
"name": "pack_gqa",
|
| 729 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 730 |
+
"default": "None",
|
| 731 |
+
"annotation": null
|
| 732 |
+
},
|
| 733 |
+
{
|
| 734 |
+
"name": "sm_margin",
|
| 735 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 736 |
+
"default": "0",
|
| 737 |
+
"annotation": null
|
| 738 |
+
}
|
| 739 |
+
],
|
| 740 |
+
"return_annotation": null
|
| 741 |
+
}
|
| 742 |
+
}
|
| 743 |
+
],
|
| 744 |
+
"layers": []
|
| 745 |
+
}
|
build/torch-stable-abi29-cu130-x86_64-linux/_flash_attn3_cuda_fe2bb66.abi3.so
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:2e59f23d082fd669404489789162e4f6dd56689862898a28c6b463361f8bd544
|
| 3 |
+
size 549647192
|
build/torch-stable-abi29-cu130-x86_64-linux/_ops.py
CHANGED
|
@@ -1,9 +1,9 @@
|
|
| 1 |
import torch
|
| 2 |
-
from . import
|
| 3 |
-
ops = torch.ops.
|
| 4 |
|
| 5 |
def add_op_namespace_prefix(op_name: str):
|
| 6 |
"""
|
| 7 |
Prefix op by namespace.
|
| 8 |
"""
|
| 9 |
-
return f"
|
|
|
|
| 1 |
import torch
|
| 2 |
+
from . import _flash_attn3_cuda_fe2bb66
|
| 3 |
+
ops = torch.ops._flash_attn3_cuda_fe2bb66
|
| 4 |
|
| 5 |
def add_op_namespace_prefix(op_name: str):
|
| 6 |
"""
|
| 7 |
Prefix op by namespace.
|
| 8 |
"""
|
| 9 |
+
return f"_flash_attn3_cuda_fe2bb66::{op_name}"
|
build/torch-stable-abi29-cu130-x86_64-linux/metadata.json
CHANGED
|
@@ -1,6 +1,6 @@
|
|
| 1 |
{
|
| 2 |
"name": "flash-attn3",
|
| 3 |
-
"id": "
|
| 4 |
"version": 2,
|
| 5 |
"kernels-minver": "0.14.0",
|
| 6 |
"license": "BSD-3-Clause",
|
|
@@ -17,20 +17,21 @@
|
|
| 17 |
"algorithm": "sha256",
|
| 18 |
"files": {
|
| 19 |
"__init__.py": "b/45TnpaDDuTwp84ILd3Q+pQCIk2RnHtSSZKRq9Lj5E=",
|
| 20 |
-
"
|
| 21 |
-
"_ops.py": "/
|
| 22 |
"flash_attn_config.py": "uxo+eyMcDit//8YDaS4Rtk5BTMKzTSK9tOKTjdWVe/w=",
|
| 23 |
-
"flash_attn_interface.py": "y0bD2JYGMFgX6iDre3zJN2sBPuq/12N6Mo55Z+HQCNQ="
|
|
|
|
| 24 |
}
|
| 25 |
},
|
| 26 |
"provenance": {
|
| 27 |
"kernel-builder": {
|
| 28 |
-
"version": "0.
|
| 29 |
-
"commit": "
|
| 30 |
"dirty": false
|
| 31 |
},
|
| 32 |
"kernel": {
|
| 33 |
-
"commit": "
|
| 34 |
"dirty": false
|
| 35 |
}
|
| 36 |
}
|
|
|
|
| 1 |
{
|
| 2 |
"name": "flash-attn3",
|
| 3 |
+
"id": "_flash_attn3_cuda_fe2bb66",
|
| 4 |
"version": 2,
|
| 5 |
"kernels-minver": "0.14.0",
|
| 6 |
"license": "BSD-3-Clause",
|
|
|
|
| 17 |
"algorithm": "sha256",
|
| 18 |
"files": {
|
| 19 |
"__init__.py": "b/45TnpaDDuTwp84ILd3Q+pQCIk2RnHtSSZKRq9Lj5E=",
|
| 20 |
+
"_flash_attn3_cuda_fe2bb66.abi3.so": "LlnyPQgv1mlARIl4kWLk9t1WaJhiiYooxrRjNh+L1UQ=",
|
| 21 |
+
"_ops.py": "qE76YSn3lYonVMHeyk13xZljOBdqQRapvPt15D/lvSI=",
|
| 22 |
"flash_attn_config.py": "uxo+eyMcDit//8YDaS4Rtk5BTMKzTSK9tOKTjdWVe/w=",
|
| 23 |
+
"flash_attn_interface.py": "y0bD2JYGMFgX6iDre3zJN2sBPuq/12N6Mo55Z+HQCNQ=",
|
| 24 |
+
"symbols.json": "dGYI9GazLj56q85OhdJTeGBnq6b9RzaJ911NQsrsjhs="
|
| 25 |
}
|
| 26 |
},
|
| 27 |
"provenance": {
|
| 28 |
"kernel-builder": {
|
| 29 |
+
"version": "0.18.0-dev0",
|
| 30 |
+
"commit": "d054906a13710b811b96697b4635f03020842733",
|
| 31 |
"dirty": false
|
| 32 |
},
|
| 33 |
"kernel": {
|
| 34 |
+
"commit": "fe2bb66fee6c2a77eccf08a0508dc57428d488ed",
|
| 35 |
"dirty": false
|
| 36 |
}
|
| 37 |
}
|
build/torch-stable-abi29-cu130-x86_64-linux/metadata.json.sigstore
CHANGED
|
@@ -1 +1 @@
|
|
| 1 |
-
{"mediaType":"application/vnd.dev.sigstore.bundle.v0.3+json", "verificationMaterial":{"certificate":{"rawBytes":"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"}, "tlogEntries":[{"logIndex":"3036566644", "logId":{"keyId":"wNI9atQGlz+VWfO6LRygH4QUfY/8W4RFwiT5i5WRgB0="}, "kindVersion":{"kind":"hashedrekord", "version":"0.0.1"}, "integratedTime":"1790868317", "inclusionPromise":{"signedEntryTimestamp":"MEUCIQDlaQVtQxXne98G9DAE5wLYD1pCUmCUQCROk5PrdcjQdwIgYimuEqA4fdTkfmQhifsFzXsjctMucpll3dGB7rjLt5k="}, "inclusionProof":{"logIndex":"2914662382", "rootHash":"gWGAJAc9iVaLBJCAy7aW32MmPea9G63I06ixoNK/yzA=", "treeSize":"2914662435", "hashes":["xDkjRoe+9NE/YBH4Em7gxsbSv/v7eQI6kkx2fOr7Vgo=", "N1HMuCBeSUXEX0eMIIQW1ojm9gKIwX+3ABL6L2CcFWM=", "NV429jvZSxlKNOQKSzF1D5STFsNHJLj4gAL0xQervnc=", "npeBYWzRzSq8gQbEA5yfhukO/qjAyuQ9PzFu1LhYphI=", "VfLY2ZTPDItNOLHK6gCW9H0J20rYCPfcNnMvN0jbjOE=", "wK5OsH9J5Hw8PFZR3xx3Q3GdTNpLn+FqMQ+r66udqF0=", "zzUPyokODCqxQJuxWw//Ch5xbzy2KJrUyyBwdJJwhLU=", "sHVmIHg06ZsXMGrd9Eo+xPXE11qUqhwMAhS2PRxcTMg=", "l9/8gARzPMRH879BPgP/QBFVPeQFIhrl5RC0/Y5wcHA=", "hEm0xTgGs7ZDCl0SvEGhPspuyoZnjSMaTu1F6ScAcH4=", "WXt1IcXPA49H96t9Z3ItfyqY4Z6slXsSWdqhfhpKl+Y=", "EsvEf22X+PURyeeu58OWCA3R8FrKbV5Rmteju3SWGuY=", "lTLoVRjXSYVbDKJgTo23loiYc3RqJdtv1AgmoVyDKlc=", "7Tzd3tvRwcVZ82zfO1MExclwm0m1ZsS2AThT/K2Tlhg=", "x+oKrxjE+hawZze3Rc5uQ+ELyYoPBRbq8pWnKfLxvPw=", "EkYZ1Q2C2GzLyxAWt8wicxm9bAs4SIiWCFwJIkXBGBw=", "W4cMs02ycm0mIMwybrVXQENsvdY9LCPKwWAnSZsKd9E=", "Xu3SV3sfB3Jk5ygvrtnnDCdDb068HxGQBt+eg1oZ+Ts=", "hjmVlWMHU8ThyhWdZrPdgN2SCOrL/5LOxhgwDX0GFr4=", "cIz1OY2tO8cVEgp0TmAffPApps2hJuAcAN+pESrpHHY=", "4lYk7cX8PszZLoyQMWjMDB1BOBn4f7hehtFiHo92e5Y=", "ttpxd0TJprOUtb/8Wf5DeIzwNtUw72Q0YnXE2LFMIZo=", "qxzHanAzz57SDdmJe0B7bJK72NTIBbwEMGBBKvDOROw=", "xH/DCseLHr9eKoYT8qsORZK7zVdEGYWHuVtsVrD95wY="], "checkpoint":{"envelope":"rekor.sigstore.dev - 1193050959916656506\n2914662435\ngWGAJAc9iVaLBJCAy7aW32MmPea9G63I06ixoNK/yzA=\n\n— rekor.sigstore.dev wNI9ajBFAiEAko9gzgzbgIWc5l4UyKkqDp27XCr0gPC7j09b4NF++JMCIB4SG/rGTgPtuu1fEOVqhknmTGK/Lfr/gAOZi8moXTd9\n"}}, "canonicalizedBody":"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"}], "timestampVerificationData":{"rfc3161Timestamps":[{"signedTimestamp":"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"}]}}, "messageSignature":{"messageDigest":{"algorithm":"SHA2_256", "digest":"ZAklDAP3I/QSZp2EgZD6tkXYQWFsz0g5dIh5rwW7SkY="}, "signature":"MEUCIDl+6462dcA30zZJa+RLZY3OJJu3q+ZwgF5x8Y01V2B2AiEAoFmJeWNo+AlXOQTmPi4JbDQqGfgjUowMWvdpN6/Hnf0="}}
|
|
|
|
| 1 |
+
{"mediaType":"application/vnd.dev.sigstore.bundle.v0.3+json", "verificationMaterial":{"certificate":{"rawBytes":"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"}, "tlogEntries":[{"logIndex":"3322705211", "logId":{"keyId":"wNI9atQGlz+VWfO6LRygH4QUfY/8W4RFwiT5i5WRgB0="}, "kindVersion":{"kind":"hashedrekord", "version":"0.0.1"}, "integratedTime":"1792423929", "inclusionPromise":{"signedEntryTimestamp":"MEYCIQCQsIJUS3Y0Yk6UDCOnAqU+iQduErWk1E+6b5ssnUw1CwIhAK/72qnOZVII+rS/BKKG24tLWCAd4RPTh7s/dnkrvP30"}, "inclusionProof":{"logIndex":"3200800949", "rootHash":"lXfsuLF6pAXluiyjGfZ70KUag+G2XJkFFJvmmnv6HnA=", "treeSize":"3200801213", "hashes":["8PjpE9lMs+UwWZAXOBbwQMqGsYMIAeFfMUsO8lkvjag=", "+AmoKSENS8BNvkOM45zoUMv+8NY/zFOimyHb92R9l3g=", "7eftHzoxb75TPJgpPZ0MG7tFPUnLy7kngswNVfOPLPs=", "HjTX6sDFzVXc4qDRRCToqf/nlZ7QGKWyZU7TfsqCy1g=", "C+idcTk5kOECt9mGNK4IefxokexI93WQAWqEKiqC0mo=", "s5Xz6U3VEx4+vp1R9fz/dp9kdz5goBtSmCpH3kqG4Ws=", "h3PoeCipxoBIC68GvrycsEroqjO2C859SfUHh6yNkgM=", "AfMFNMBs7MKZ7wrUGPusBcmSr5ytzSGtoWojae6bFN4=", "b1QxgPy+ptzj4NfnAzTkklUq6+X5Q8J+7FXNQGAcJKU=", "cmAt0qje3Kd/EjUJnYTX38NLOtqyGSFOtx2g8hxibeI=", "11lr1PAKx8EF0MQ8eyTG2Okak7VNiBdR4LabBugomSw=", "uGEIH4dmv632MUSdl5AVDq3lFz9EPX5Y4BnnOA6WyoM=", "D37E7MwfvuENRJ4BUsqOTbzKKFjz+Dkh4kKr7xMQyU8=", "7WFpoaiQCzLwM3o2YI5Ngo8J7JD0OfyCCQtmXhWxVes=", "nyQV/+cWQUG5Hb2PKjQQNkQIJmz2qzAZKBQlkHGIknM=", "G+nk7mgE5h6dUmv2UY7NDJrY3WvrZY31raaO1AETt+o=", "qlTWNDzERnTBQVcHHXXVPjiUP0SsikGxfcxVlgqVFDo=", "gg9GNApNke3+zNhl4tsWtmIycCFrOMCB3lys6XuSQNQ=", "NlOclldmKPA1tMKIskqHAwuQDYHp31YwpK+OOMrMccc=", "qxzHanAzz57SDdmJe0B7bJK72NTIBbwEMGBBKvDOROw=", "xH/DCseLHr9eKoYT8qsORZK7zVdEGYWHuVtsVrD95wY="], "checkpoint":{"envelope":"rekor.sigstore.dev - 1193050959916656506\n3200801213\nlXfsuLF6pAXluiyjGfZ70KUag+G2XJkFFJvmmnv6HnA=\n\n— rekor.sigstore.dev wNI9ajBEAiAmEXrn24SzhF2D3En/7ItrG5THmHS60qt9fWCND3avxQIgSNCB+vvInFinuSFRnt/t+PlzBXC3wVHl0bECXDBSIIw=\n"}}, "canonicalizedBody":"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"}], "timestampVerificationData":{"rfc3161Timestamps":[{"signedTimestamp":"MIICyzADAgEAMIICwgYJKoZIhvcNAQcCoIICszCCAq8CAQMxDTALBglghkgBZQMEAgEwgbgGCyqGSIb3DQEJEAEEoIGoBIGlMIGiAgEBBgkrBgEEAYO/MAIwMTANBglghkgBZQMEAgEFAAQguCQuSbjrO+A68ljA7EQ+KFcqGe24QNQBQISi2NvAbZcCFQD69t6c7kt5iV7Uxjn8lO60b08OtxgPMjAyNjEwMTkxNTMyMDhaMAMCAQGgMqQwMC4xFTATBgNVBAoTDHNpZ3N0b3JlLmRldjEVMBMGA1UEAxMMc2lnc3RvcmUtdHNhoAAxggHcMIIB2AIBATBRMDkxFTATBgNVBAoTDHNpZ3N0b3JlLmRldjEgMB4GA1UEAxMXc2lnc3RvcmUtdHNhLXNlbGZzaWduZWQCFDoTVC8MkGHuvMFDL8uKjosqI4sMMAsGCWCGSAFlAwQCAaCB/DAaBgkqhkiG9w0BCQMxDQYLKoZIhvcNAQkQAQQwHAYJKoZIhvcNAQkFMQ8XDTI2MTAxOTE1MzIwOFowLwYJKoZIhvcNAQkEMSIEIC91TVNYweHW3ShrHf4wi0oceD33DFYfh+UuF/Pq23ZwMIGOBgsqhkiG9w0BCRACLzF/MH0wezB5BCCF+Se8B6tiysO0Q1bBDvyBssaIP9p6uebYcNnROs0FtzBVMD2kOzA5MRUwEwYDVQQKEwxzaWdzdG9yZS5kZXYxIDAeBgNVBAMTF3NpZ3N0b3JlLXRzYS1zZWxmc2lnbmVkAhQ6E1QvDJBh7rzBQy/Lio6LKiOLDDAKBggqhkjOPQQDAgRoMGYCMQDY7VoA5jnYKPpzOgQHA6K3B2NB4ZfkcXGRTPcQ64GhF7eu6nvOZENB+mMbZI2e5r8CMQCgMOmVtLx6e9VWOOo+kWEGJHJRBRmxLALfDfQMONj2m4OEEUGDoFN6OK2fJeW/tQk="}]}}, "messageSignature":{"messageDigest":{"algorithm":"SHA2_256", "digest":"+u7+5pg8x/tji2ku4gCtLnMcX/VrnZJGmSmcg7cpgsY="}, "signature":"MEUCIQDqbzmJ2RrhxX1Oam2VsgRLon2Q9EyCGXbanyridoZFwQIgZ649dIYHCP/aoA51nsbI3clUX6gELNA7j8U9RMNS3Ug="}}
|
build/torch-stable-abi29-cu130-x86_64-linux/symbols.json
ADDED
|
@@ -0,0 +1,745 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"schema_version": 1,
|
| 3 |
+
"module": "_flash_attn3_cuda_fe2bb66",
|
| 4 |
+
"functions": [
|
| 5 |
+
{
|
| 6 |
+
"name": "add_op_namespace_prefix",
|
| 7 |
+
"qualname": "add_op_namespace_prefix",
|
| 8 |
+
"module": "_flash_attn3_cuda_fe2bb66._ops",
|
| 9 |
+
"kind": "function",
|
| 10 |
+
"doc": "Prefix op by namespace.",
|
| 11 |
+
"signature": {
|
| 12 |
+
"parameters": [
|
| 13 |
+
{
|
| 14 |
+
"name": "op_name",
|
| 15 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 16 |
+
"default": null,
|
| 17 |
+
"annotation": "str"
|
| 18 |
+
}
|
| 19 |
+
],
|
| 20 |
+
"return_annotation": null
|
| 21 |
+
}
|
| 22 |
+
},
|
| 23 |
+
{
|
| 24 |
+
"name": "flash_attn_combine",
|
| 25 |
+
"qualname": "flash_attn_combine",
|
| 26 |
+
"module": "_flash_attn3_cuda_fe2bb66.flash_attn_interface",
|
| 27 |
+
"kind": "function",
|
| 28 |
+
"doc": null,
|
| 29 |
+
"signature": {
|
| 30 |
+
"parameters": [
|
| 31 |
+
{
|
| 32 |
+
"name": "out_partial",
|
| 33 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 34 |
+
"default": null,
|
| 35 |
+
"annotation": null
|
| 36 |
+
},
|
| 37 |
+
{
|
| 38 |
+
"name": "lse_partial",
|
| 39 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 40 |
+
"default": null,
|
| 41 |
+
"annotation": null
|
| 42 |
+
},
|
| 43 |
+
{
|
| 44 |
+
"name": "out",
|
| 45 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 46 |
+
"default": "None",
|
| 47 |
+
"annotation": null
|
| 48 |
+
},
|
| 49 |
+
{
|
| 50 |
+
"name": "out_dtype",
|
| 51 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 52 |
+
"default": "None",
|
| 53 |
+
"annotation": null
|
| 54 |
+
}
|
| 55 |
+
],
|
| 56 |
+
"return_annotation": null
|
| 57 |
+
}
|
| 58 |
+
},
|
| 59 |
+
{
|
| 60 |
+
"name": "flash_attn_func",
|
| 61 |
+
"qualname": "flash_attn_func",
|
| 62 |
+
"module": "_flash_attn3_cuda_fe2bb66.flash_attn_interface",
|
| 63 |
+
"kind": "function",
|
| 64 |
+
"doc": "dropout_p should be set to 0.0 during evaluation\nSupports multi-query and grouped-query attention (MQA/GQA) by passing in KV with fewer heads\nthan Q. Note that the number of heads in Q must be divisible by the number of heads in KV.\nFor example, if Q has 6 heads and K, V have 2 heads, head 0, 1, 2 of Q will attention to head\n0 of K, V, and head 3, 4, 5 of Q will attention to head 1 of K, V.\n\nIf causal=True, the causal mask is aligned to the bottom right corner of the attention matrix.\nFor example, if seqlen_q = 2 and seqlen_k = 5, the causal mask (1 = keep, 0 = masked out) is:\n 1 1 1 1 0\n 1 1 1 1 1\nIf seqlen_q = 5 and seqlen_k = 2, the causal mask is:\n 0 0\n 0 0\n 0 0\n 1 0\n 1 1\nIf the row of the mask is all zero, the output will be zero.\n\nIf window_size != (-1, -1), implements sliding window local attention. Query at position i\nwill only attend to keys between\n[i + seqlen_k - seqlen_q - window_size[0], i + seqlen_k - seqlen_q + window_size[1]] inclusive.\n\nArguments:\n q: (batch_size, seqlen, nheads, headdim)\n k: (batch_size, seqlen, nheads_k, headdim)\n v: (batch_size, seqlen, nheads_k, headdim)\n dropout_p: float. Dropout probability.\n softmax_scale: float. The scaling of QK^T before applying softmax.\n Default to 1 / sqrt(headdim).\n causal: bool. Whether to apply causal attention mask (e.g., for auto-regressive modeling).\n window_size: (left, right). If not (-1, -1), implements sliding window local attention.\n alibi_slopes: (nheads,) or (batch_size, nheads), fp32. A bias of\n (-alibi_slope * |i + seqlen_k - seqlen_q - j|)\n is added to the attention score of query i and key j.\n deterministic: bool. Whether to use the deterministic implementation of the backward pass,\n which is slightly slower and uses more memory. The forward pass is always deterministic.\n return_attn_probs: bool. Whether to return the attention probabilities. This option is for\n testing only. The returned probabilities are not guaranteed to be correct\n (they might not have the right scaling).\nReturn:\n out: (batch_size, seqlen, nheads, headdim).\n softmax_lse [optional, if return_attn_probs=True]: (batch_size, nheads, seqlen). The\n logsumexp of each row of the matrix QK^T * scaling (e.g., log of the softmax\n normalization factor).",
|
| 65 |
+
"signature": {
|
| 66 |
+
"parameters": [
|
| 67 |
+
{
|
| 68 |
+
"name": "q",
|
| 69 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 70 |
+
"default": null,
|
| 71 |
+
"annotation": null
|
| 72 |
+
},
|
| 73 |
+
{
|
| 74 |
+
"name": "k",
|
| 75 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 76 |
+
"default": null,
|
| 77 |
+
"annotation": null
|
| 78 |
+
},
|
| 79 |
+
{
|
| 80 |
+
"name": "v",
|
| 81 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 82 |
+
"default": null,
|
| 83 |
+
"annotation": null
|
| 84 |
+
},
|
| 85 |
+
{
|
| 86 |
+
"name": "softmax_scale",
|
| 87 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 88 |
+
"default": "None",
|
| 89 |
+
"annotation": null
|
| 90 |
+
},
|
| 91 |
+
{
|
| 92 |
+
"name": "causal",
|
| 93 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 94 |
+
"default": "False",
|
| 95 |
+
"annotation": null
|
| 96 |
+
},
|
| 97 |
+
{
|
| 98 |
+
"name": "qv",
|
| 99 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 100 |
+
"default": "None",
|
| 101 |
+
"annotation": null
|
| 102 |
+
},
|
| 103 |
+
{
|
| 104 |
+
"name": "q_descale",
|
| 105 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 106 |
+
"default": "None",
|
| 107 |
+
"annotation": null
|
| 108 |
+
},
|
| 109 |
+
{
|
| 110 |
+
"name": "k_descale",
|
| 111 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 112 |
+
"default": "None",
|
| 113 |
+
"annotation": null
|
| 114 |
+
},
|
| 115 |
+
{
|
| 116 |
+
"name": "v_descale",
|
| 117 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 118 |
+
"default": "None",
|
| 119 |
+
"annotation": null
|
| 120 |
+
},
|
| 121 |
+
{
|
| 122 |
+
"name": "window_size",
|
| 123 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 124 |
+
"default": "(-1, -1)",
|
| 125 |
+
"annotation": null
|
| 126 |
+
},
|
| 127 |
+
{
|
| 128 |
+
"name": "attention_chunk",
|
| 129 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 130 |
+
"default": "0",
|
| 131 |
+
"annotation": null
|
| 132 |
+
},
|
| 133 |
+
{
|
| 134 |
+
"name": "softcap",
|
| 135 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 136 |
+
"default": "0.0",
|
| 137 |
+
"annotation": null
|
| 138 |
+
},
|
| 139 |
+
{
|
| 140 |
+
"name": "num_splits",
|
| 141 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 142 |
+
"default": "1",
|
| 143 |
+
"annotation": null
|
| 144 |
+
},
|
| 145 |
+
{
|
| 146 |
+
"name": "pack_gqa",
|
| 147 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 148 |
+
"default": "None",
|
| 149 |
+
"annotation": null
|
| 150 |
+
},
|
| 151 |
+
{
|
| 152 |
+
"name": "deterministic",
|
| 153 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 154 |
+
"default": "False",
|
| 155 |
+
"annotation": null
|
| 156 |
+
},
|
| 157 |
+
{
|
| 158 |
+
"name": "sm_margin",
|
| 159 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 160 |
+
"default": "0",
|
| 161 |
+
"annotation": null
|
| 162 |
+
},
|
| 163 |
+
{
|
| 164 |
+
"name": "return_attn_probs",
|
| 165 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 166 |
+
"default": "False",
|
| 167 |
+
"annotation": null
|
| 168 |
+
}
|
| 169 |
+
],
|
| 170 |
+
"return_annotation": null
|
| 171 |
+
}
|
| 172 |
+
},
|
| 173 |
+
{
|
| 174 |
+
"name": "flash_attn_qkvpacked_func",
|
| 175 |
+
"qualname": "flash_attn_qkvpacked_func",
|
| 176 |
+
"module": "_flash_attn3_cuda_fe2bb66.flash_attn_interface",
|
| 177 |
+
"kind": "function",
|
| 178 |
+
"doc": "dropout_p should be set to 0.0 during evaluation\nIf Q, K, V are already stacked into 1 tensor, this function will be faster than\ncalling flash_attn_func on Q, K, V since the backward pass avoids explicit concatenation\nof the gradients of Q, K, V.\nFor multi-query and grouped-query attention (MQA/GQA), please see\nflash_attn_kvpacked_func and flash_attn_func.\n\nIf window_size != (-1, -1), implements sliding window local attention. Query at position i\nwill only attend to keys between [i - window_size[0], i + window_size[1]] inclusive.\n\nArguments:\n qkv: (batch_size, seqlen, 3, nheads, headdim)\n dropout_p: float. Dropout probability.\n softmax_scale: float. The scaling of QK^T before applying softmax.\n Default to 1 / sqrt(headdim).\n causal: bool. Whether to apply causal attention mask (e.g., for auto-regressive modeling).\n window_size: (left, right). If not (-1, -1), implements sliding window local attention.\n softcap: float. Anything > 0 activates softcapping attention.\n alibi_slopes: (nheads,) or (batch_size, nheads), fp32. A bias of (-alibi_slope * |i - j|) is added to\n the attention score of query i and key j.\n deterministic: bool. Whether to use the deterministic implementation of the backward pass,\n which is slightly slower and uses more memory. The forward pass is always deterministic.\n return_attn_probs: bool. Whether to return the attention probabilities. This option is for\n testing only. The returned probabilities are not guaranteed to be correct\n (they might not have the right scaling).\nReturn:\n out: (batch_size, seqlen, nheads, headdim).\n softmax_lse [optional, if return_attn_probs=True]: (batch_size, nheads, seqlen). The\n logsumexp of each row of the matrix QK^T * scaling (e.g., log of the softmax\n normalization factor).\n S_dmask [optional, if return_attn_probs=True]: (batch_size, nheads, seqlen, seqlen).\n The output of softmax (possibly with different scaling). It also encodes the dropout\n pattern (negative means that location was dropped, nonnegative means it was kept).",
|
| 179 |
+
"signature": {
|
| 180 |
+
"parameters": [
|
| 181 |
+
{
|
| 182 |
+
"name": "qkv",
|
| 183 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 184 |
+
"default": null,
|
| 185 |
+
"annotation": null
|
| 186 |
+
},
|
| 187 |
+
{
|
| 188 |
+
"name": "softmax_scale",
|
| 189 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 190 |
+
"default": "None",
|
| 191 |
+
"annotation": null
|
| 192 |
+
},
|
| 193 |
+
{
|
| 194 |
+
"name": "causal",
|
| 195 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 196 |
+
"default": "False",
|
| 197 |
+
"annotation": null
|
| 198 |
+
},
|
| 199 |
+
{
|
| 200 |
+
"name": "q_descale",
|
| 201 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 202 |
+
"default": "None",
|
| 203 |
+
"annotation": null
|
| 204 |
+
},
|
| 205 |
+
{
|
| 206 |
+
"name": "k_descale",
|
| 207 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 208 |
+
"default": "None",
|
| 209 |
+
"annotation": null
|
| 210 |
+
},
|
| 211 |
+
{
|
| 212 |
+
"name": "v_descale",
|
| 213 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 214 |
+
"default": "None",
|
| 215 |
+
"annotation": null
|
| 216 |
+
},
|
| 217 |
+
{
|
| 218 |
+
"name": "window_size",
|
| 219 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 220 |
+
"default": "(-1, -1)",
|
| 221 |
+
"annotation": null
|
| 222 |
+
},
|
| 223 |
+
{
|
| 224 |
+
"name": "attention_chunk",
|
| 225 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 226 |
+
"default": "0",
|
| 227 |
+
"annotation": null
|
| 228 |
+
},
|
| 229 |
+
{
|
| 230 |
+
"name": "softcap",
|
| 231 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 232 |
+
"default": "0.0",
|
| 233 |
+
"annotation": null
|
| 234 |
+
},
|
| 235 |
+
{
|
| 236 |
+
"name": "deterministic",
|
| 237 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 238 |
+
"default": "False",
|
| 239 |
+
"annotation": null
|
| 240 |
+
},
|
| 241 |
+
{
|
| 242 |
+
"name": "num_heads_q",
|
| 243 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 244 |
+
"default": "None",
|
| 245 |
+
"annotation": null
|
| 246 |
+
},
|
| 247 |
+
{
|
| 248 |
+
"name": "sm_margin",
|
| 249 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 250 |
+
"default": "0",
|
| 251 |
+
"annotation": null
|
| 252 |
+
},
|
| 253 |
+
{
|
| 254 |
+
"name": "return_attn_probs",
|
| 255 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 256 |
+
"default": "False",
|
| 257 |
+
"annotation": null
|
| 258 |
+
}
|
| 259 |
+
],
|
| 260 |
+
"return_annotation": null
|
| 261 |
+
}
|
| 262 |
+
},
|
| 263 |
+
{
|
| 264 |
+
"name": "flash_attn_varlen_func",
|
| 265 |
+
"qualname": "flash_attn_varlen_func",
|
| 266 |
+
"module": "_flash_attn3_cuda_fe2bb66.flash_attn_interface",
|
| 267 |
+
"kind": "function",
|
| 268 |
+
"doc": null,
|
| 269 |
+
"signature": {
|
| 270 |
+
"parameters": [
|
| 271 |
+
{
|
| 272 |
+
"name": "q",
|
| 273 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 274 |
+
"default": null,
|
| 275 |
+
"annotation": null
|
| 276 |
+
},
|
| 277 |
+
{
|
| 278 |
+
"name": "k",
|
| 279 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 280 |
+
"default": null,
|
| 281 |
+
"annotation": null
|
| 282 |
+
},
|
| 283 |
+
{
|
| 284 |
+
"name": "v",
|
| 285 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 286 |
+
"default": null,
|
| 287 |
+
"annotation": null
|
| 288 |
+
},
|
| 289 |
+
{
|
| 290 |
+
"name": "cu_seqlens_q",
|
| 291 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 292 |
+
"default": null,
|
| 293 |
+
"annotation": null
|
| 294 |
+
},
|
| 295 |
+
{
|
| 296 |
+
"name": "cu_seqlens_k",
|
| 297 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 298 |
+
"default": null,
|
| 299 |
+
"annotation": null
|
| 300 |
+
},
|
| 301 |
+
{
|
| 302 |
+
"name": "max_seqlen_q",
|
| 303 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 304 |
+
"default": null,
|
| 305 |
+
"annotation": null
|
| 306 |
+
},
|
| 307 |
+
{
|
| 308 |
+
"name": "max_seqlen_k",
|
| 309 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 310 |
+
"default": null,
|
| 311 |
+
"annotation": null
|
| 312 |
+
},
|
| 313 |
+
{
|
| 314 |
+
"name": "seqused_q",
|
| 315 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 316 |
+
"default": "None",
|
| 317 |
+
"annotation": null
|
| 318 |
+
},
|
| 319 |
+
{
|
| 320 |
+
"name": "seqused_k",
|
| 321 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 322 |
+
"default": "None",
|
| 323 |
+
"annotation": null
|
| 324 |
+
},
|
| 325 |
+
{
|
| 326 |
+
"name": "softmax_scale",
|
| 327 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 328 |
+
"default": "None",
|
| 329 |
+
"annotation": null
|
| 330 |
+
},
|
| 331 |
+
{
|
| 332 |
+
"name": "causal",
|
| 333 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 334 |
+
"default": "False",
|
| 335 |
+
"annotation": null
|
| 336 |
+
},
|
| 337 |
+
{
|
| 338 |
+
"name": "qv",
|
| 339 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 340 |
+
"default": "None",
|
| 341 |
+
"annotation": null
|
| 342 |
+
},
|
| 343 |
+
{
|
| 344 |
+
"name": "q_descale",
|
| 345 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 346 |
+
"default": "None",
|
| 347 |
+
"annotation": null
|
| 348 |
+
},
|
| 349 |
+
{
|
| 350 |
+
"name": "k_descale",
|
| 351 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 352 |
+
"default": "None",
|
| 353 |
+
"annotation": null
|
| 354 |
+
},
|
| 355 |
+
{
|
| 356 |
+
"name": "v_descale",
|
| 357 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 358 |
+
"default": "None",
|
| 359 |
+
"annotation": null
|
| 360 |
+
},
|
| 361 |
+
{
|
| 362 |
+
"name": "window_size",
|
| 363 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 364 |
+
"default": "(-1, -1)",
|
| 365 |
+
"annotation": null
|
| 366 |
+
},
|
| 367 |
+
{
|
| 368 |
+
"name": "attention_chunk",
|
| 369 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 370 |
+
"default": "0",
|
| 371 |
+
"annotation": null
|
| 372 |
+
},
|
| 373 |
+
{
|
| 374 |
+
"name": "softcap",
|
| 375 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 376 |
+
"default": "0.0",
|
| 377 |
+
"annotation": null
|
| 378 |
+
},
|
| 379 |
+
{
|
| 380 |
+
"name": "num_splits",
|
| 381 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 382 |
+
"default": "1",
|
| 383 |
+
"annotation": null
|
| 384 |
+
},
|
| 385 |
+
{
|
| 386 |
+
"name": "pack_gqa",
|
| 387 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 388 |
+
"default": "None",
|
| 389 |
+
"annotation": null
|
| 390 |
+
},
|
| 391 |
+
{
|
| 392 |
+
"name": "deterministic",
|
| 393 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 394 |
+
"default": "False",
|
| 395 |
+
"annotation": null
|
| 396 |
+
},
|
| 397 |
+
{
|
| 398 |
+
"name": "sm_margin",
|
| 399 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 400 |
+
"default": "0",
|
| 401 |
+
"annotation": null
|
| 402 |
+
},
|
| 403 |
+
{
|
| 404 |
+
"name": "return_attn_probs",
|
| 405 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 406 |
+
"default": "False",
|
| 407 |
+
"annotation": null
|
| 408 |
+
}
|
| 409 |
+
],
|
| 410 |
+
"return_annotation": null
|
| 411 |
+
}
|
| 412 |
+
},
|
| 413 |
+
{
|
| 414 |
+
"name": "flash_attn_with_kvcache",
|
| 415 |
+
"qualname": "flash_attn_with_kvcache",
|
| 416 |
+
"module": "_flash_attn3_cuda_fe2bb66.flash_attn_interface",
|
| 417 |
+
"kind": "function",
|
| 418 |
+
"doc": "If k and v are not None, k_cache and v_cache will be updated *inplace* with the new values from\nk and v. This is useful for incremental decoding: you can pass in the cached keys/values from\nthe previous step, and update them with the new keys/values from the current step, and do\nattention with the updated cache, all in 1 kernel.\n\nIf you pass in k / v, you must make sure that the cache is large enough to hold the new values.\nFor example, the KV cache could be pre-allocated with the max sequence length, and you can use\ncache_seqlens to keep track of the current sequence lengths of each sequence in the batch.\n\nAlso apply rotary embedding if rotary_cos and rotary_sin are passed in. The key @k will be\nrotated by rotary_cos and rotary_sin at indices cache_seqlens, cache_seqlens + 1, etc.\nIf causal or local (i.e., window_size != (-1, -1)), the query @q will be rotated by rotary_cos\nand rotary_sin at indices cache_seqlens, cache_seqlens + 1, etc.\nIf not causal and not local, the query @q will be rotated by rotary_cos and rotary_sin at\nindices cache_seqlens only (i.e. we consider all tokens in @q to be at position cache_seqlens).\n\nSee tests/test_flash_attn.py::test_flash_attn_kvcache for examples of how to use this function.\n\nSupports multi-query and grouped-query attention (MQA/GQA) by passing in KV with fewer heads\nthan Q. Note that the number of heads in Q must be divisible by the number of heads in KV.\nFor example, if Q has 6 heads and K, V have 2 heads, head 0, 1, 2 of Q will attention to head\n0 of K, V, and head 3, 4, 5 of Q will attention to head 1 of K, V.\n\nIf causal=True, the causal mask is aligned to the bottom right corner of the attention matrix.\nFor example, if seqlen_q = 2 and seqlen_k = 5, the causal mask (1 = keep, 0 = masked out) is:\n 1 1 1 1 0\n 1 1 1 1 1\nIf seqlen_q = 5 and seqlen_k = 2, the causal mask is:\n 0 0\n 0 0\n 0 0\n 1 0\n 1 1\nIf the row of the mask is all zero, the output will be zero.\n\nIf window_size != (-1, -1), implements sliding window local attention. Query at position i\nwill only attend to keys between\n[i + seqlen_k - seqlen_q - window_size[0], i + seqlen_k - seqlen_q + window_size[1]] inclusive.\n\nNote: Does not support backward pass.\n\nArguments:\n q: (batch_size, seqlen, nheads, headdim)\n k_cache: (batch_size_cache, seqlen_cache, nheads_k, headdim) if there's no page_table,\n or (num_blocks, page_block_size, nheads_k, headdim) if there's a page_table (i.e. paged KV cache)\n page_block_size can be arbitrary (e.g, 1, 2, 3, 64, etc.).\n v_cache: (batch_size_cache, seqlen_cache, nheads_k, headdim_v) if there's no page_table,\n or (num_blocks, page_block_size, nheads_k, headdim_v) if there's a page_table (i.e. paged KV cache)\n k [optional]: (batch_size, seqlen_new, nheads_k, headdim). If not None, we concatenate\n k with k_cache, starting at the indices specified by cache_seqlens.\n v [optional]: (batch_size, seqlen_new, nheads_k, headdim_v). Similar to k.\n qv [optional]: (batch_size, seqlen, nheads, headdim_v)\n rotary_cos [optional]: (seqlen_ro, rotary_dim / 2). If not None, we apply rotary embedding\n to k and q. Only applicable if k and v are passed in. rotary_dim must be divisible by 16.\n rotary_sin [optional]: (seqlen_ro, rotary_dim / 2). Similar to rotary_cos.\n cache_seqlens: int, or (batch_size,), dtype torch.int32. The sequence lengths of the\n KV cache.\n cache_batch_idx: (batch_size,), dtype torch.int32. The indices used to index into the KV cache.\n If None, we assume that the batch indices are [0, 1, 2, ..., batch_size - 1].\n If the indices are not distinct, and k and v are provided, the values updated in the cache\n might come from any of the duplicate indices.\n cache_leftpad: (batch_size,), dtype torch.int32. The index that the KV cache starts. If None, assume 0.\n page_table [optional]: (batch_size, max_num_blocks_per_seq), dtype torch.int32.\n softmax_scale: float. The scaling of QK^T before applying softmax.\n Default to 1 / sqrt(headdim).\n causal: bool. Whether to apply causal attention mask (e.g., for auto-regressive modeling).\n window_size: (left, right). If not (-1, -1), implements sliding window local attention.\n softcap: float. Anything > 0 activates softcapping attention.\n rotary_interleaved: bool. Only applicable if rotary_cos and rotary_sin are passed in.\n If True, rotary embedding will combine dimensions 0 & 1, 2 & 3, etc. If False,\n rotary embedding will combine dimensions 0 & rotary_dim / 2, 1 & rotary_dim / 2 + 1\n (i.e. GPT-NeoX style).\n num_splits: int. If > 1, split the key/value into this many chunks along the sequence.\n If num_splits == 1, we don't split the key/value. If num_splits == 0, we use a heuristic\n to automatically determine the number of splits.\n Don't change this unless you know what you are doing.\n return_softmax_lse: bool. Whether to return the logsumexp of the attention scores.\n\nReturn:\n out: (batch_size, seqlen, nheads, headdim).\n softmax_lse [optional, if return_softmax_lse=True]: (batch_size, nheads, seqlen). The\n logsumexp of each row of the matrix QK^T * scaling (e.g., log of the softmax\n normalization factor).",
|
| 419 |
+
"signature": {
|
| 420 |
+
"parameters": [
|
| 421 |
+
{
|
| 422 |
+
"name": "q",
|
| 423 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 424 |
+
"default": null,
|
| 425 |
+
"annotation": null
|
| 426 |
+
},
|
| 427 |
+
{
|
| 428 |
+
"name": "k_cache",
|
| 429 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 430 |
+
"default": null,
|
| 431 |
+
"annotation": null
|
| 432 |
+
},
|
| 433 |
+
{
|
| 434 |
+
"name": "v_cache",
|
| 435 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 436 |
+
"default": null,
|
| 437 |
+
"annotation": null
|
| 438 |
+
},
|
| 439 |
+
{
|
| 440 |
+
"name": "k",
|
| 441 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 442 |
+
"default": "None",
|
| 443 |
+
"annotation": null
|
| 444 |
+
},
|
| 445 |
+
{
|
| 446 |
+
"name": "v",
|
| 447 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 448 |
+
"default": "None",
|
| 449 |
+
"annotation": null
|
| 450 |
+
},
|
| 451 |
+
{
|
| 452 |
+
"name": "qv",
|
| 453 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 454 |
+
"default": "None",
|
| 455 |
+
"annotation": null
|
| 456 |
+
},
|
| 457 |
+
{
|
| 458 |
+
"name": "rotary_cos",
|
| 459 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 460 |
+
"default": "None",
|
| 461 |
+
"annotation": null
|
| 462 |
+
},
|
| 463 |
+
{
|
| 464 |
+
"name": "rotary_sin",
|
| 465 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 466 |
+
"default": "None",
|
| 467 |
+
"annotation": null
|
| 468 |
+
},
|
| 469 |
+
{
|
| 470 |
+
"name": "cache_seqlens",
|
| 471 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 472 |
+
"default": "None",
|
| 473 |
+
"annotation": "int | torch.Tensor | None"
|
| 474 |
+
},
|
| 475 |
+
{
|
| 476 |
+
"name": "cache_batch_idx",
|
| 477 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 478 |
+
"default": "None",
|
| 479 |
+
"annotation": "torch.Tensor | None"
|
| 480 |
+
},
|
| 481 |
+
{
|
| 482 |
+
"name": "cache_leftpad",
|
| 483 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 484 |
+
"default": "None",
|
| 485 |
+
"annotation": "torch.Tensor | None"
|
| 486 |
+
},
|
| 487 |
+
{
|
| 488 |
+
"name": "page_table",
|
| 489 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 490 |
+
"default": "None",
|
| 491 |
+
"annotation": "torch.Tensor | None"
|
| 492 |
+
},
|
| 493 |
+
{
|
| 494 |
+
"name": "cu_seqlens_q",
|
| 495 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 496 |
+
"default": "None",
|
| 497 |
+
"annotation": "torch.Tensor | None"
|
| 498 |
+
},
|
| 499 |
+
{
|
| 500 |
+
"name": "cu_seqlens_k_new",
|
| 501 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 502 |
+
"default": "None",
|
| 503 |
+
"annotation": "torch.Tensor | None"
|
| 504 |
+
},
|
| 505 |
+
{
|
| 506 |
+
"name": "max_seqlen_q",
|
| 507 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 508 |
+
"default": "None",
|
| 509 |
+
"annotation": "int | None"
|
| 510 |
+
},
|
| 511 |
+
{
|
| 512 |
+
"name": "rotary_seqlens",
|
| 513 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 514 |
+
"default": "None",
|
| 515 |
+
"annotation": "torch.Tensor | None"
|
| 516 |
+
},
|
| 517 |
+
{
|
| 518 |
+
"name": "q_descale",
|
| 519 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 520 |
+
"default": "None",
|
| 521 |
+
"annotation": "torch.Tensor | None"
|
| 522 |
+
},
|
| 523 |
+
{
|
| 524 |
+
"name": "k_descale",
|
| 525 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 526 |
+
"default": "None",
|
| 527 |
+
"annotation": "torch.Tensor | None"
|
| 528 |
+
},
|
| 529 |
+
{
|
| 530 |
+
"name": "v_descale",
|
| 531 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 532 |
+
"default": "None",
|
| 533 |
+
"annotation": "torch.Tensor | None"
|
| 534 |
+
},
|
| 535 |
+
{
|
| 536 |
+
"name": "softmax_scale",
|
| 537 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 538 |
+
"default": "None",
|
| 539 |
+
"annotation": null
|
| 540 |
+
},
|
| 541 |
+
{
|
| 542 |
+
"name": "causal",
|
| 543 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 544 |
+
"default": "False",
|
| 545 |
+
"annotation": null
|
| 546 |
+
},
|
| 547 |
+
{
|
| 548 |
+
"name": "window_size",
|
| 549 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 550 |
+
"default": "(-1, -1)",
|
| 551 |
+
"annotation": null
|
| 552 |
+
},
|
| 553 |
+
{
|
| 554 |
+
"name": "attention_chunk",
|
| 555 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 556 |
+
"default": "0",
|
| 557 |
+
"annotation": null
|
| 558 |
+
},
|
| 559 |
+
{
|
| 560 |
+
"name": "softcap",
|
| 561 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 562 |
+
"default": "0.0",
|
| 563 |
+
"annotation": null
|
| 564 |
+
},
|
| 565 |
+
{
|
| 566 |
+
"name": "rotary_interleaved",
|
| 567 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 568 |
+
"default": "True",
|
| 569 |
+
"annotation": null
|
| 570 |
+
},
|
| 571 |
+
{
|
| 572 |
+
"name": "scheduler_metadata",
|
| 573 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 574 |
+
"default": "None",
|
| 575 |
+
"annotation": null
|
| 576 |
+
},
|
| 577 |
+
{
|
| 578 |
+
"name": "num_splits",
|
| 579 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 580 |
+
"default": "0",
|
| 581 |
+
"annotation": null
|
| 582 |
+
},
|
| 583 |
+
{
|
| 584 |
+
"name": "pack_gqa",
|
| 585 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 586 |
+
"default": "None",
|
| 587 |
+
"annotation": null
|
| 588 |
+
},
|
| 589 |
+
{
|
| 590 |
+
"name": "sm_margin",
|
| 591 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 592 |
+
"default": "0",
|
| 593 |
+
"annotation": null
|
| 594 |
+
},
|
| 595 |
+
{
|
| 596 |
+
"name": "return_softmax_lse",
|
| 597 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 598 |
+
"default": "False",
|
| 599 |
+
"annotation": null
|
| 600 |
+
}
|
| 601 |
+
],
|
| 602 |
+
"return_annotation": null
|
| 603 |
+
}
|
| 604 |
+
},
|
| 605 |
+
{
|
| 606 |
+
"name": "get_scheduler_metadata",
|
| 607 |
+
"qualname": "get_scheduler_metadata",
|
| 608 |
+
"module": "_flash_attn3_cuda_fe2bb66.flash_attn_interface",
|
| 609 |
+
"kind": "function",
|
| 610 |
+
"doc": null,
|
| 611 |
+
"signature": {
|
| 612 |
+
"parameters": [
|
| 613 |
+
{
|
| 614 |
+
"name": "batch_size",
|
| 615 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 616 |
+
"default": null,
|
| 617 |
+
"annotation": null
|
| 618 |
+
},
|
| 619 |
+
{
|
| 620 |
+
"name": "max_seqlen_q",
|
| 621 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 622 |
+
"default": null,
|
| 623 |
+
"annotation": null
|
| 624 |
+
},
|
| 625 |
+
{
|
| 626 |
+
"name": "max_seqlen_k",
|
| 627 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 628 |
+
"default": null,
|
| 629 |
+
"annotation": null
|
| 630 |
+
},
|
| 631 |
+
{
|
| 632 |
+
"name": "num_heads_q",
|
| 633 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 634 |
+
"default": null,
|
| 635 |
+
"annotation": null
|
| 636 |
+
},
|
| 637 |
+
{
|
| 638 |
+
"name": "num_heads_kv",
|
| 639 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 640 |
+
"default": null,
|
| 641 |
+
"annotation": null
|
| 642 |
+
},
|
| 643 |
+
{
|
| 644 |
+
"name": "headdim",
|
| 645 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 646 |
+
"default": null,
|
| 647 |
+
"annotation": null
|
| 648 |
+
},
|
| 649 |
+
{
|
| 650 |
+
"name": "cache_seqlens",
|
| 651 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 652 |
+
"default": null,
|
| 653 |
+
"annotation": "torch.Tensor"
|
| 654 |
+
},
|
| 655 |
+
{
|
| 656 |
+
"name": "qkv_dtype",
|
| 657 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 658 |
+
"default": "torch.bfloat16",
|
| 659 |
+
"annotation": null
|
| 660 |
+
},
|
| 661 |
+
{
|
| 662 |
+
"name": "headdim_v",
|
| 663 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 664 |
+
"default": "None",
|
| 665 |
+
"annotation": null
|
| 666 |
+
},
|
| 667 |
+
{
|
| 668 |
+
"name": "cu_seqlens_q",
|
| 669 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 670 |
+
"default": "None",
|
| 671 |
+
"annotation": "torch.Tensor | None"
|
| 672 |
+
},
|
| 673 |
+
{
|
| 674 |
+
"name": "cu_seqlens_k_new",
|
| 675 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 676 |
+
"default": "None",
|
| 677 |
+
"annotation": "torch.Tensor | None"
|
| 678 |
+
},
|
| 679 |
+
{
|
| 680 |
+
"name": "cache_leftpad",
|
| 681 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 682 |
+
"default": "None",
|
| 683 |
+
"annotation": "torch.Tensor | None"
|
| 684 |
+
},
|
| 685 |
+
{
|
| 686 |
+
"name": "page_size",
|
| 687 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 688 |
+
"default": "None",
|
| 689 |
+
"annotation": "int | None"
|
| 690 |
+
},
|
| 691 |
+
{
|
| 692 |
+
"name": "max_seqlen_k_new",
|
| 693 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 694 |
+
"default": "0",
|
| 695 |
+
"annotation": null
|
| 696 |
+
},
|
| 697 |
+
{
|
| 698 |
+
"name": "causal",
|
| 699 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 700 |
+
"default": "False",
|
| 701 |
+
"annotation": null
|
| 702 |
+
},
|
| 703 |
+
{
|
| 704 |
+
"name": "window_size",
|
| 705 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 706 |
+
"default": "(-1, -1)",
|
| 707 |
+
"annotation": null
|
| 708 |
+
},
|
| 709 |
+
{
|
| 710 |
+
"name": "attention_chunk",
|
| 711 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 712 |
+
"default": "0",
|
| 713 |
+
"annotation": null
|
| 714 |
+
},
|
| 715 |
+
{
|
| 716 |
+
"name": "has_softcap",
|
| 717 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 718 |
+
"default": "False",
|
| 719 |
+
"annotation": null
|
| 720 |
+
},
|
| 721 |
+
{
|
| 722 |
+
"name": "num_splits",
|
| 723 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 724 |
+
"default": "0",
|
| 725 |
+
"annotation": null
|
| 726 |
+
},
|
| 727 |
+
{
|
| 728 |
+
"name": "pack_gqa",
|
| 729 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 730 |
+
"default": "None",
|
| 731 |
+
"annotation": null
|
| 732 |
+
},
|
| 733 |
+
{
|
| 734 |
+
"name": "sm_margin",
|
| 735 |
+
"kind": "POSITIONAL_OR_KEYWORD",
|
| 736 |
+
"default": "0",
|
| 737 |
+
"annotation": null
|
| 738 |
+
}
|
| 739 |
+
],
|
| 740 |
+
"return_annotation": null
|
| 741 |
+
}
|
| 742 |
+
}
|
| 743 |
+
],
|
| 744 |
+
"layers": []
|
| 745 |
+
}
|