Kernels:
Trusted publisher
Uploaded using `kernel-builder`.
Browse files- build/torch-stable-abi29-cu126-x86_64-linux/{_flash_attn3_cuda_5fe9dbb.abi3.so → _flash_attn3_cuda_1ac140b.abi3.so} +1 -1
- build/torch-stable-abi29-cu126-x86_64-linux/_ops.py +3 -3
- build/torch-stable-abi29-cu126-x86_64-linux/metadata.json +7 -7
- build/torch-stable-abi29-cu126-x86_64-linux/metadata.json.sigstore +1 -1
- build/torch-stable-abi29-cu126-x86_64-linux/symbols.json +8 -8
- build/torch-stable-abi29-cu128-x86_64-linux/{_flash_attn3_cuda_5fe9dbb.abi3.so → _flash_attn3_cuda_1ac140b.abi3.so} +1 -1
- build/torch-stable-abi29-cu128-x86_64-linux/_ops.py +3 -3
- build/torch-stable-abi29-cu128-x86_64-linux/metadata.json +7 -7
- build/torch-stable-abi29-cu128-x86_64-linux/metadata.json.sigstore +1 -1
- build/torch-stable-abi29-cu128-x86_64-linux/symbols.json +8 -8
- build/torch-stable-abi29-cu130-x86_64-linux/{_flash_attn3_cuda_5fe9dbb.abi3.so → _flash_attn3_cuda_1ac140b.abi3.so} +1 -1
- build/torch-stable-abi29-cu130-x86_64-linux/_ops.py +3 -3
- build/torch-stable-abi29-cu130-x86_64-linux/metadata.json +7 -7
- build/torch-stable-abi29-cu130-x86_64-linux/metadata.json.sigstore +1 -1
- build/torch-stable-abi29-cu130-x86_64-linux/symbols.json +8 -8
- build/torch-stable-abi29-cu132-x86_64-linux/__init__.py +20 -0
- build/torch-stable-abi29-cu132-x86_64-linux/_flash_attn3_cuda_1ac140b.abi3.so +3 -0
- build/torch-stable-abi29-cu132-x86_64-linux/_ops.py +9 -0
- build/torch-stable-abi29-cu132-x86_64-linux/flash_attn_config.py +7 -0
- build/torch-stable-abi29-cu132-x86_64-linux/flash_attn_interface.py +1127 -0
- build/torch-stable-abi29-cu132-x86_64-linux/metadata.json +38 -0
- build/torch-stable-abi29-cu132-x86_64-linux/metadata.json.sigstore +1 -0
- build/torch-stable-abi29-cu132-x86_64-linux/symbols.json +745 -0
- build/torch-stable-abi29-cu134-x86_64-linux/{_flash_attn3_cuda_5fe9dbb.abi3.so → _flash_attn3_cuda_1ac140b.abi3.so} +1 -1
- build/torch-stable-abi29-cu134-x86_64-linux/_ops.py +3 -3
- build/torch-stable-abi29-cu134-x86_64-linux/metadata.json +7 -7
- build/torch-stable-abi29-cu134-x86_64-linux/metadata.json.sigstore +1 -1
- build/torch-stable-abi29-cu134-x86_64-linux/symbols.json +8 -8
build/torch-stable-abi29-cu126-x86_64-linux/{_flash_attn3_cuda_5fe9dbb.abi3.so → _flash_attn3_cuda_1ac140b.abi3.so}
RENAMED
|
@@ -1,3 +1,3 @@
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
-
oid sha256:
|
| 3 |
size 526430928
|
|
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:ff9e54b5b75c870a36881f13ab094466a0d171b620f1f3153900cc8db397e7bb
|
| 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_1ac140b
|
| 3 |
+
ops = torch.ops._flash_attn3_cuda_1ac140b
|
| 4 |
|
| 5 |
def add_op_namespace_prefix(op_name: str):
|
| 6 |
"""
|
| 7 |
Prefix op by namespace.
|
| 8 |
"""
|
| 9 |
+
return f"_flash_attn3_cuda_1ac140b::{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",
|
|
@@ -17,21 +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 |
-
"symbols.json": "
|
| 25 |
}
|
| 26 |
},
|
| 27 |
"provenance": {
|
| 28 |
"kernel-builder": {
|
| 29 |
-
"version": "0.
|
| 30 |
-
"commit": "
|
| 31 |
"dirty": false
|
| 32 |
},
|
| 33 |
"kernel": {
|
| 34 |
-
"commit": "
|
| 35 |
"dirty": false
|
| 36 |
}
|
| 37 |
}
|
|
|
|
| 1 |
{
|
| 2 |
"name": "flash-attn3",
|
| 3 |
+
"id": "_flash_attn3_cuda_1ac140b",
|
| 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_1ac140b.abi3.so": "/55Utbdchwo2iB8TqwlEZqDRcbYg8fMVOQDMjbOX57s=",
|
| 21 |
+
"_ops.py": "NH6/V6V8u1HinwVJvVwoG30z34Aa9OL/sgYZocQi7a8=",
|
| 22 |
"flash_attn_config.py": "uxo+eyMcDit//8YDaS4Rtk5BTMKzTSK9tOKTjdWVe/w=",
|
| 23 |
"flash_attn_interface.py": "y0bD2JYGMFgX6iDre3zJN2sBPuq/12N6Mo55Z+HQCNQ=",
|
| 24 |
+
"symbols.json": "qzII4ieoZRHwN7dHbNFGa995mlTpWyJs1X5eRdNLU5o="
|
| 25 |
}
|
| 26 |
},
|
| 27 |
"provenance": {
|
| 28 |
"kernel-builder": {
|
| 29 |
+
"version": "0.19.0-dev0",
|
| 30 |
+
"commit": "88e9dd07eb4e639b85799c14b2d8af14a0d9580a",
|
| 31 |
"dirty": false
|
| 32 |
},
|
| 33 |
"kernel": {
|
| 34 |
+
"commit": "1ac140bf36af3c8ff0e73262add55fb8df939e26",
|
| 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":"3601916445", "logId":{"keyId":"wNI9atQGlz+VWfO6LRygH4QUfY/8W4RFwiT5i5WRgB0="}, "kindVersion":{"kind":"hashedrekord", "version":"0.0.1"}, "integratedTime":"1793392059", "inclusionPromise":{"signedEntryTimestamp":"MEYCIQCqvjRw7iWZHyvGYDTNGjhHZVmBnDnrjtbPsTiBtdCyDgIhAJ4Tch8KIDUvQoyPjIEPpLYbP855/bu2jicOleoutgtq"}, "inclusionProof":{"logIndex":"3480012183", "rootHash":"Ro7OiKO9pRdsKIMVACkDBth9yE8Q5k5+0HvQnGP0coQ=", "treeSize":"3480013432", "hashes":["8PZyV514+G3213Md6+9TOIw4We04CkNslCtlb9ckEj8=", "3zwbsopnGhSGgf6/jruuYtgkmahMlI4TPCM11KKn5y8=", "37teURBIKWibJg+7usJzQvqBHiRVwyAQFLaZlDZu0dc=", "bR+KcqiWHz0x/BxUChPBwOmT2morm4Toa1TBg6jksfI=", "bKE1GX067r4YXaIcd8y/5nLlWVhGqzoWNWrLJ8h9bls=", "sxLt4bfGOo0GD2tazozkWEUgh+cgwBHI6dpRzRUIG10=", "kQIMtL5EbwGmFlAaVHzqPwDHYH1qBFxjfmmYjO3lXWU=", "tJ3h3oNTKitRMUd/oFYSrW6icPwykE6T6CLo+pBAA1k=", "OvOURmw0F9uwldegFdRp/pXLKqXkZ3LGbQmAewamy8E=", "f6UAPcoerwIoewj/j5cecUpexTLUvMD1qU38eIkAuQU=", "dW/l8RkQdmdaDT9baNKoEVPgEKgfcyUxVX3vCpjgdc4=", "DuGe1ieTCpe0LDFqXfq6YOF8tPsigOSKbBAwG5IHr7w=", "FGzETXd85KSZDiS854wzNAsJ+POTz2es7GVf9cwUo+s=", "oRoie0t/H+0ZyKRjI6KiDz9stRDDR7EJu7t8qrlVT0Y=", "a8cmhOLBStNzSIVybWzlBd1UxGct2Vy2Eggt5g2B26w=", "39Z7bKKSjPtMpTU8xt8EybgTkCg1qk72XWkQqE5WRSU=", "63Z7Fze4Sm0atCwhCb6EG/N+w2qm+Q+7+dWxf2+uPzA=", "njdy0pCf5uBXEpRPrJGvuVRc6ibqank2dtzoCQ0SR18=", "e3OaWwzL1tEWYXiUzGpH+1v9CiFPwweyfRzIh3Z+uO4=", "b+938HIdfjdD/+LgIs3AVRE3Wc65/1NSSAL4OJ5FUXE=", "kQz/4ND7LdhfAQn+idNsO+ZFC32sfkYRlhjWF5W+sPc=", "zGfeoWKmrZlXMOvyi3R2GDbGpmI1skJW/wLTXCCgF0o=", "1DaDKzrCNNBhmSdY5vFjz+NpAmxmYfYEgBCdPLrMP1A=", "xH/DCseLHr9eKoYT8qsORZK7zVdEGYWHuVtsVrD95wY="], "checkpoint":{"envelope":"rekor.sigstore.dev - 1193050959916656506\n3480013432\nRo7OiKO9pRdsKIMVACkDBth9yE8Q5k5+0HvQnGP0coQ=\n\n— rekor.sigstore.dev wNI9ajBFAiEAxQViIXmGZasPb3PyhHfk9Crb81C2a2qGVEEftxVOeDsCIAKMs0PpWyr4aFmK568X+a7uYdx8lwd4+MTuyHZriVoi\n"}}, "canonicalizedBody":"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"}], "timestampVerificationData":{"rfc3161Timestamps":[{"signedTimestamp":"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"}]}}, "messageSignature":{"messageDigest":{"algorithm":"SHA2_256", "digest":"/8B6G+jbUGkDEEgK7QxyZHf9LVW2qlL0lerTaw7t2q0="}, "signature":"MEQCIEwWGkl5LQ2zR8rGecXSFe4liko6zH6rWpRZfQZUfF/kAiAD3vR4uDgRpTrjo8kb5mXMe4xymExyX6q8fPM/Q/H01Q=="}}
|
|
|
|
| 1 |
+
{"mediaType":"application/vnd.dev.sigstore.bundle.v0.3+json", "verificationMaterial":{"certificate":{"rawBytes":"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"}, "tlogEntries":[{"logIndex":"3654775158", "logId":{"keyId":"wNI9atQGlz+VWfO6LRygH4QUfY/8W4RFwiT5i5WRgB0="}, "kindVersion":{"kind":"hashedrekord", "version":"0.0.1"}, "integratedTime":"1793669004", "inclusionPromise":{"signedEntryTimestamp":"MEQCIFO5ibMxJDgIsb3oy3yEeXBg63i6Fu652Il5NUIw4hJCAiBMf2T+EnBi1yCLBc8zaPws2CKpSbpMwMu3hZtGoxTgNg=="}, "inclusionProof":{"logIndex":"3532870896", "rootHash":"xzXRErS1Ssn61ZOrFb0nQEGeKbR+q1ErmxhJANvKQQk=", "treeSize":"3532871865", "hashes":["qQwqx5B6JFaihitf4ICEXXIGOTx5+PZm2OTX5T59Dxc=", "lSGgLcMVMW8XDjEOTuY9KJfW3jTAFgrMvJpjUMHanZw=", "DSOKs6/LWXHF3VgBrkfKftqzToGa6adWQwSuqP3tfso=", "/PIQ+sN62A+pasz8aV9ZprTk/KMEoq4NmgrnP8CTV6E=", "B1UIbj2LtZpANKi0+6XXYUTq7OgReOfsMcH5A8d7yck=", "I411IwzgG6q+krnaUrY05MO4nIBnQ69F4UPXsRzVI7I=", "FS7mEicmJr6srHd07h1f9Iq7m2xpGumCHm/iMPIdxNM=", "BEUo+6BxNhRmj37f+aTTiU0wEn5Y8Sa7FzGXKQ2kdQY=", "tD3qlgdDMlVP2E5K9/TbodqZFN9lHDh01fMoZqghCbk=", "3NoeN70cMfe4tSqKFRRqAFNzZQ64rMmPMjbW6FidztE=", "lwMS7BnOTD+Tip+5bAgFoEoTLmhBNuEUNrbKlb5r1Fc=", "x30FICvbcqPmxyauu1pwEl0iA62LRIG3P46P22VV/+M=", "l2F8XDZWSpFsRQNOEJbYK/sHi0FHC7rqAIaBXYhU1pQ=", "jXRnHZhdIEDN/dOSOGnahiPHiCL8svXXZkNlpklSSDE=", "JgdHnyIJAAm3PA7mCUWsdLQ/ZPPYPC6Oh1Gw79WHtpg=", "MlqLoiFm8Oo5bbmftEQbgcSKXfa+i5FCiungGzOC1uo=", "QYcxedNInCRxBCNHkv8RT2T790mUNX70dZsQ78gLIYQ=", "ogjRZ88pnEW/ANOWeLXfhYL/XdgeGJy0zJTgCayJlGE=", "PA9MhoiExYPp/vXh8tVCBrATkoY31KStgze+rovDTkA=", "SlixnS34BhAO+yRI2FaGPSOh9V6JSVNPfLl1bHxrcQY=", "1DaDKzrCNNBhmSdY5vFjz+NpAmxmYfYEgBCdPLrMP1A=", "xH/DCseLHr9eKoYT8qsORZK7zVdEGYWHuVtsVrD95wY="], "checkpoint":{"envelope":"rekor.sigstore.dev - 1193050959916656506\n3532871865\nxzXRErS1Ssn61ZOrFb0nQEGeKbR+q1ErmxhJANvKQQk=\n\n— rekor.sigstore.dev wNI9ajBEAiAszY8b8WntKsd5mxWwbd4fGQ9nmR6E5iO60raupECLPwIgSofgCfCWH+EFnnjzPy6PN7pGqTzPeNDyhlQ2nQaaitg=\n"}}, "canonicalizedBody":"eyJhcGlWZXJzaW9uIjoiMC4wLjEiLCJraW5kIjoiaGFzaGVkcmVrb3JkIiwic3BlYyI6eyJkYXRhIjp7Imhhc2giOnsiYWxnb3JpdGhtIjoic2hhMjU2IiwidmFsdWUiOiI3ODQ0YTI5NjVkNjQ0NzE3NTMwYzcyMzljMmY5NmFiMzBiNmYzYTE4YWM0M2M0MTJlMTI3Y2M0OGEwZDc5YzY2In19LCJzaWduYXR1cmUiOnsiY29udGVudCI6Ik1FWUNJUUNNK0x0MUErZnpCTTlBbUNDcnZFUW1QazRPcDJBcFNZMGxrSTdTeDVja21nSWhBSkpSL1FqenNyeWNwNjdETjhaV3dBMTRTQzFIeWVMY3hPVnBEOVFyMnVFQSIsInB1YmxpY0tleSI6eyJjb250ZW50IjoiTFMwdExTMUNSVWRKVGlCRFJWSlVTVVpKUTBGVVJTMHRMUzB0Q2sxSlNVaExla05EUW5KSFowRjNTVUpCWjBsVlJVcHFjbkF4VTBKdlRqZENWVEpwYldReU16RlRiblpCVkU5UmQwTm5XVWxMYjFwSmVtb3dSVUYzVFhjS1RucEZWazFDVFVkQk1WVkZRMmhOVFdNeWJHNWpNMUoyWTIxVmRWcEhWakpOVWpSM1NFRlpSRlpSVVVSRmVGWjZZVmRrZW1SSE9YbGFVekZ3WW01U2JBcGpiVEZzV2tkc2FHUkhWWGRJYUdOT1RXcFplRTFVUVhwTlJFVjVUWHBCTlZkb1kwNU5hbGw0VFZSQmVrMUVSWHBOZWtFMVYycEJRVTFHYTNkRmQxbElDa3R2V2tsNmFqQkRRVkZaU1V0dldrbDZhakJFUVZGalJGRm5RVVZTUldscmNYUk5ObFo0YWxSWVRHdzJXVTFZVnpOTVRHRk5WbkJOWkRGVlZGZEdRM1VLTkVGUE5qVXJUQ3RqVEVJNGVXbFFaRlZIVEN0QlpTOUpkMlpDVlVWVmJtOTRSWE40ZGtoUWFYRXpjaXQyYW5sNVYzRlBRMEprUVhkbloxaE5UVUUwUndwQk1WVmtSSGRGUWk5M1VVVkJkMGxJWjBSQlZFSm5UbFpJVTFWRlJFUkJTMEpuWjNKQ1owVkdRbEZqUkVGNlFXWkNaMDVXU0ZOTlJVZEVRVmRuUWxSbUNqQXJibEJXYVZGU2JIWnRiekpQYTI5V1lVeEhUR2hvYTFCNlFuSkNaMDVXU0ZKRlFrRm1PRVZaVkVKbWFHd3hiMlJJVW5kamVtOTJUREprY0dSSGFERUtXV2sxYW1JeU1IWmhTRlp1V2pKc2RWb3lXbWhaTWxWMllUSldlV0p0Vm5OamVURnFZakl4ZEdSWE5YQmtTR3QyVEcxa2NHUkhhREZaYVRrellqTktjZ3BhYlhoMlpETk5kbGx1Vm5CaVIxRjFaVmRHZEdKRlFubGFWMXA2VERKb2JGbFhVbnBNTWpGb1lWYzBkMDlSV1V0TGQxbENRa0ZIUkhaNlFVSkJVVkZ5Q21GSVVqQmpTRTAyVEhrNU1HSXlkR3hpYVRWb1dUTlNjR0l5TlhwTWJXUndaRWRvTVZsdVZucGFXRXBxWWpJMU1GcFhOVEJNYlU1MllsUkJaa0puYjNJS1FtZEZSVUZaVHk5TlFVVkRRa0pHTTJJelNuSmFiWGgyWkRFNWEyRllUbmRaV0ZKcVlVUkJNa0puYjNKQ1owVkZRVmxQTDAxQlJVUkNRMmQ0V1ZkTmVBcE9SRUpwV21wTk1sbFhXWHBaZW1odFdtcENiRTU2VFhsT2FrcG9Xa2RSTVU1WFdtbFBSMUp0VDFSTk5WcFVTVEpOUWsxSFEybHpSMEZSVVVKbk56aDNDa0ZSVVVWQ1ZVb3hZVmQ0YTAxRGMwZERhWE5IUVZGUlFtYzNPSGRCVVZWRlNGZG9NVm95WkhCaWJXUnRXVmRPYkV3eWRHeGpiVFZzWWtoTmRGa3lPWFFLWWxoV2RXRllValZOUWpCSFEybHpSMEZSVVVKbk56aDNRVkZaUlVRelNteGFiazEyWVVkV2FGcElUWFppVjBad1ltcEJOMEpuYjNKQ1owVkZRVmxQTHdwTlFVVkpRa013VFVzeWFEQmtTRUo2VDJrNGRtUkhPWEphVnpSMVdWZE9NR0ZYT1hWamVUVnVZVmhTYjJSWFNqRmpNbFo1V1RJNWRXUkhWblZrUXpWcUNtSXlNSGRpVVZsTFMzZFpRa0pCUjBSMmVrRkNRMUZTWmtSR01XOWtTRkozWTNwdmRrd3laSEJrUjJneFdXazFhbUl5TUhaaFNGWnVXakpzZFZveVdtZ0tXVEpWZG1FeVZubGliVlp6WTNreGFtSXlNWFJrVnpWd1pFaHJka3h0WkhCa1IyZ3hXV2s1TTJJelNuSmFiWGgyWkROTmRsbHVWbkJpUjFGMVpWZEdkQXBpUlVKNVdsZGFla3d5YUd4WlYxSjZUREl4YUdGWE5IZFBRVmxMUzNkWlFrSkJSMFIyZWtGQ1EyZFJjVVJEWjNoWlYwMTRUa1JDYVZwcVRUSlpWMWw2Q2xsNmFHMWFha0pzVG5wTmVVNXFTbWhhUjFFeFRsZGFhVTlIVW0xUFZFMDFXbFJKTWsxQ2MwZERhWE5IUVZGUlFtYzNPSGRCVVhORlJGRjNUR015Vm5NS1dta3hiMkl6VGpCYVYxRjNVVUZaUzB0M1dVSkNRVWRFZG5wQlFrUkJVWGxFUkVKdlpFaFNkMk42YjNaTU1tUndaRWRvTVZscE5XcGlNakIyWVVoV2JncGFNbXgxV2pKYWFGa3lWWFpoTWxaNVltMVdjMk41TVdwaU1qRjBaRmMxY0dSSWEzZFBRVmxMUzNkWlFrSkJSMFIyZWtGQ1JGRlJjVVJEWjNoWlYwMTRDazVFUW1sYWFrMHlXVmRaZWxsNmFHMWFha0pzVG5wTmVVNXFTbWhhUjFFeFRsZGFhVTlIVW0xUFZFMDFXbFJKTWsxQ09FZERhWE5IUVZGUlFtYzNPSGNLUVZFMFJVVlJkMUJqYlZadFkzazViMXBYUm10amVUbDBXVmRzZFUxQ2IwZERhWE5IUVZGUlFtYzNPSGRCVVRoRlJFRjNTMDFVUVROTlZGRXpUbFJWZVFwUFZFRjFRbWR2Y2tKblJVVkJXVTh2VFVGRlVVSkRRVTFJYldnd1pFaENlazlwT0haYU1td3dZVWhXYVV4dFRuWmlVemx2WkZka2JtRlhOVzVhYlVacUNscFVRVmxDWjI5eVFtZEZSVUZaVHk5TlFVVlNRa0Z2VFVORVNURk9la2wzVG5wUmVrMUhNRWREYVhOSFFWRlJRbWMzT0hkQlVrbEZXSGQ0WkdGSVVqQUtZMGhOTmt4NU9XNWhXRkp2WkZkSmRWa3lPWFJNTW1neFdqSmtjR0p0WkcxWlYwNXNUREowYkdOdE5XeGlTRTEwV1RJNWRHSllWblZoV0ZJMVRIazFiZ3BoV0ZKdlpGZEpkbVF5T1hsaE1scHpZak5rZWt3eVNqRmhWM2hyVEc1c2FHSlhlRUZqYlZadFkzazViMXBYUm10amVUbDBXVmRzZFUxRVowZERhWE5IQ2tGUlVVSm5OemgzUVZKTlJVdG5kMjlOVjBacVRWUlJkMWx0V1hwT2JVWnRUVEpOTkZwdFdYZGFWR042VFdwWmVWbFhVbXRPVkZadFdXcG9hMXBxYTNvS1QxZFZlVTVxUVdoQ1oyOXlRbWRGUlVGWlR5OU5RVVZWUWtKTlRVVllaSFpqYlhSdFlrYzVNMWd5VW5Cak0wSm9aRWRPYjAxSFVVZERhWE5IUVZGUlFncG5OemgzUVZKVlJWWm5lRlZoU0ZJd1kwaE5Oa3g1T1c1aFdGSnZaRmRKZFZreU9YUk1NbWd4V2pKa2NHSnRaRzFaVjA1c1RESjBiR050Tld4aVNFMTBDbGt5T1hSaVdGWjFZVmhTTlV3eVJtcGtSMngyWW01TmRtTnVWblZqZVRnd1RWUlZkMDlVWnpOT2VsRXhUV2s1YUdSSVVteGlXRUl3WTNrNGVFMUNXVWNLUTJselIwRlJVVUpuTnpoM1FWSlpSVU5CZDBkalNGWnBZa2RzYWsxRldVZERhWE5IUVZGUlFtYzNPSGRCVW1kRlQwRjNNbU50Vm5kaWVuQnZaRmRrYmdwaFZ6VnVXbTFHYWxwVE9YSmFXRXAxV2xkNGVreFhUblppVnpFeFltMXNNR1ZVY0hsYVYxazJZMjFXYldONU9XOWFWMFpyWTNrNWRGbFhiSFZOU1VkS0NrSm5iM0pDWjBWRlFXUmFOVUZuVVVOQ1NITkZaVkZDTTBGSVZVRXpWREIzWVhOaVNFVlVTbXBIVWpSamJWZGpNMEZ4U2t0WWNtcGxVRXN6TDJnMGNIa0taME00Y0Rkdk5FRkJRVWRvYm5kRVEzUlJRVUZDUVUxQlVtcENSVUZwUVhWNFN6QkNORmhSV2tKbllXaG1RbTFxYkN0SVZsVkNiRVV6V2s5R2IzUTJVUXBXZFVodFQyaEdObXhCU1dkUU9XNVVZV2hwUm10UGVXZEZhM3AzTjJrNVkxY3dWRE15WVZsc1IxUmFPVFI1VGswM1RtMTNOMjl2ZDBObldVbExiMXBKQ25wcU1FVkJkMDFFWVVGQmQxcFJTWGhCU2taVGJXMTFOMkpaTmxjclpURnJhVkpsTUVkYVIwNWhSSGhsVXpSa1EwdEVaamxJYUhZMGR6UjZRbkF4ZWs0S1VXZG5hMmN4ZDNSaVZqUndhMHB2Tkc1QlNYZEhaMWxKZHpGNmNUWXJSWEZWU1Uxd1VtVktjM2xXZFc1cVIwOHpPRTU1YURGTEswcEZlRGhhVVd0RWVBbzVTVVZJTVRrMk0wdHhRVGszVUdOUVN6WlVWQW90TFMwdExVVk9SQ0JEUlZKVVNVWkpRMEZVUlMwdExTMHRDZz09In19fX0="}], "timestampVerificationData":{"rfc3161Timestamps":[{"signedTimestamp":"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"}]}}, "messageSignature":{"messageDigest":{"algorithm":"SHA2_256", "digest":"eESill1kRxdTDHI5wvlqswtvOhisQ8QS4SfMSKDXnGY="}, "signature":"MEYCIQCM+Lt1A+fzBM9AmCCrvEQmPk4Op2ApSY0lkI7Sx5ckmgIhAJJR/Qjzsrycp67DN8ZWwA14SC1HyeLcxOVpD9Qr2uEA"}}
|
build/torch-stable-abi29-cu126-x86_64-linux/symbols.json
CHANGED
|
@@ -1,11 +1,11 @@
|
|
| 1 |
{
|
| 2 |
"schema_version": 1,
|
| 3 |
-
"module": "
|
| 4 |
"functions": [
|
| 5 |
{
|
| 6 |
"name": "add_op_namespace_prefix",
|
| 7 |
"qualname": "add_op_namespace_prefix",
|
| 8 |
-
"module": "
|
| 9 |
"kind": "function",
|
| 10 |
"doc": "Prefix op by namespace.",
|
| 11 |
"signature": {
|
|
@@ -23,7 +23,7 @@
|
|
| 23 |
{
|
| 24 |
"name": "flash_attn_combine",
|
| 25 |
"qualname": "flash_attn_combine",
|
| 26 |
-
"module": "
|
| 27 |
"kind": "function",
|
| 28 |
"doc": null,
|
| 29 |
"signature": {
|
|
@@ -59,7 +59,7 @@
|
|
| 59 |
{
|
| 60 |
"name": "flash_attn_func",
|
| 61 |
"qualname": "flash_attn_func",
|
| 62 |
-
"module": "
|
| 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": {
|
|
@@ -173,7 +173,7 @@
|
|
| 173 |
{
|
| 174 |
"name": "flash_attn_qkvpacked_func",
|
| 175 |
"qualname": "flash_attn_qkvpacked_func",
|
| 176 |
-
"module": "
|
| 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": {
|
|
@@ -263,7 +263,7 @@
|
|
| 263 |
{
|
| 264 |
"name": "flash_attn_varlen_func",
|
| 265 |
"qualname": "flash_attn_varlen_func",
|
| 266 |
-
"module": "
|
| 267 |
"kind": "function",
|
| 268 |
"doc": null,
|
| 269 |
"signature": {
|
|
@@ -413,7 +413,7 @@
|
|
| 413 |
{
|
| 414 |
"name": "flash_attn_with_kvcache",
|
| 415 |
"qualname": "flash_attn_with_kvcache",
|
| 416 |
-
"module": "
|
| 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": {
|
|
@@ -605,7 +605,7 @@
|
|
| 605 |
{
|
| 606 |
"name": "get_scheduler_metadata",
|
| 607 |
"qualname": "get_scheduler_metadata",
|
| 608 |
-
"module": "
|
| 609 |
"kind": "function",
|
| 610 |
"doc": null,
|
| 611 |
"signature": {
|
|
|
|
| 1 |
{
|
| 2 |
"schema_version": 1,
|
| 3 |
+
"module": "_flash_attn3_cuda_1ac140b",
|
| 4 |
"functions": [
|
| 5 |
{
|
| 6 |
"name": "add_op_namespace_prefix",
|
| 7 |
"qualname": "add_op_namespace_prefix",
|
| 8 |
+
"module": "_flash_attn3_cuda_1ac140b._ops",
|
| 9 |
"kind": "function",
|
| 10 |
"doc": "Prefix op by namespace.",
|
| 11 |
"signature": {
|
|
|
|
| 23 |
{
|
| 24 |
"name": "flash_attn_combine",
|
| 25 |
"qualname": "flash_attn_combine",
|
| 26 |
+
"module": "_flash_attn3_cuda_1ac140b.flash_attn_interface",
|
| 27 |
"kind": "function",
|
| 28 |
"doc": null,
|
| 29 |
"signature": {
|
|
|
|
| 59 |
{
|
| 60 |
"name": "flash_attn_func",
|
| 61 |
"qualname": "flash_attn_func",
|
| 62 |
+
"module": "_flash_attn3_cuda_1ac140b.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": {
|
|
|
|
| 173 |
{
|
| 174 |
"name": "flash_attn_qkvpacked_func",
|
| 175 |
"qualname": "flash_attn_qkvpacked_func",
|
| 176 |
+
"module": "_flash_attn3_cuda_1ac140b.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": {
|
|
|
|
| 263 |
{
|
| 264 |
"name": "flash_attn_varlen_func",
|
| 265 |
"qualname": "flash_attn_varlen_func",
|
| 266 |
+
"module": "_flash_attn3_cuda_1ac140b.flash_attn_interface",
|
| 267 |
"kind": "function",
|
| 268 |
"doc": null,
|
| 269 |
"signature": {
|
|
|
|
| 413 |
{
|
| 414 |
"name": "flash_attn_with_kvcache",
|
| 415 |
"qualname": "flash_attn_with_kvcache",
|
| 416 |
+
"module": "_flash_attn3_cuda_1ac140b.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": {
|
|
|
|
| 605 |
{
|
| 606 |
"name": "get_scheduler_metadata",
|
| 607 |
"qualname": "get_scheduler_metadata",
|
| 608 |
+
"module": "_flash_attn3_cuda_1ac140b.flash_attn_interface",
|
| 609 |
"kind": "function",
|
| 610 |
"doc": null,
|
| 611 |
"signature": {
|
build/torch-stable-abi29-cu128-x86_64-linux/{_flash_attn3_cuda_5fe9dbb.abi3.so → _flash_attn3_cuda_1ac140b.abi3.so}
RENAMED
|
@@ -1,3 +1,3 @@
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
-
oid sha256:
|
| 3 |
size 530776216
|
|
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:081450e7c15ceef7e1b263706cfc825fa527dd6a521c162b9fa8bb49e74c31f6
|
| 3 |
size 530776216
|
build/torch-stable-abi29-cu128-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_1ac140b
|
| 3 |
+
ops = torch.ops._flash_attn3_cuda_1ac140b
|
| 4 |
|
| 5 |
def add_op_namespace_prefix(op_name: str):
|
| 6 |
"""
|
| 7 |
Prefix op by namespace.
|
| 8 |
"""
|
| 9 |
+
return f"_flash_attn3_cuda_1ac140b::{op_name}"
|
build/torch-stable-abi29-cu128-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",
|
|
@@ -17,21 +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 |
-
"symbols.json": "
|
| 25 |
}
|
| 26 |
},
|
| 27 |
"provenance": {
|
| 28 |
"kernel-builder": {
|
| 29 |
-
"version": "0.
|
| 30 |
-
"commit": "
|
| 31 |
"dirty": false
|
| 32 |
},
|
| 33 |
"kernel": {
|
| 34 |
-
"commit": "
|
| 35 |
"dirty": false
|
| 36 |
}
|
| 37 |
}
|
|
|
|
| 1 |
{
|
| 2 |
"name": "flash-attn3",
|
| 3 |
+
"id": "_flash_attn3_cuda_1ac140b",
|
| 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_1ac140b.abi3.so": "CBRQ58Fc7vfhsmNwbPyCX6Un3WpSHBYrn6i7SedMMfY=",
|
| 21 |
+
"_ops.py": "NH6/V6V8u1HinwVJvVwoG30z34Aa9OL/sgYZocQi7a8=",
|
| 22 |
"flash_attn_config.py": "uxo+eyMcDit//8YDaS4Rtk5BTMKzTSK9tOKTjdWVe/w=",
|
| 23 |
"flash_attn_interface.py": "y0bD2JYGMFgX6iDre3zJN2sBPuq/12N6Mo55Z+HQCNQ=",
|
| 24 |
+
"symbols.json": "qzII4ieoZRHwN7dHbNFGa995mlTpWyJs1X5eRdNLU5o="
|
| 25 |
}
|
| 26 |
},
|
| 27 |
"provenance": {
|
| 28 |
"kernel-builder": {
|
| 29 |
+
"version": "0.19.0-dev0",
|
| 30 |
+
"commit": "88e9dd07eb4e639b85799c14b2d8af14a0d9580a",
|
| 31 |
"dirty": false
|
| 32 |
},
|
| 33 |
"kernel": {
|
| 34 |
+
"commit": "1ac140bf36af3c8ff0e73262add55fb8df939e26",
|
| 35 |
"dirty": false
|
| 36 |
}
|
| 37 |
}
|
build/torch-stable-abi29-cu128-x86_64-linux/metadata.json.sigstore
CHANGED
|
@@ -1 +1 @@
|
|
| 1 |
-
{"mediaType":"application/vnd.dev.sigstore.bundle.v0.3+json", "verificationMaterial":{"certificate":{"rawBytes":"MIIHLDCCBrGgAwIBAgIUcSc+zuCpk91awisj0GBx71VBloEwCgYIKoZIzj0EAwMwNzEVMBMGA1UEChMMc2lnc3RvcmUuZGV2MR4wHAYDVQQDExVzaWdzdG9yZS1pbnRlcm1lZGlhdGUwHhcNMjYxMDMwMjAyNzQ2WhcNMjYxMDMwMjAzNzQ2WjAAMFkwEwYHKoZIzj0CAQYIKoZIzj0DAQcDQgAEIR4V+kaNe0AAMiS0I4B97AudMBSPB9iFcV8I2RhRGM1Nj05qH21NKE0z3r5unryalm2KJQSvL+/B6f2W/mABbqOCBdAwggXMMA4GA1UdDwEB/wQEAwIHgDATBgNVHSUEDDAKBggrBgEFBQcDAzAfBgNVHSMEGDAWgBTf0+nPViQRlvmo2OkoVaLGLhhkPzBrBgNVHREBAf8EYTBfhl1odHRwczovL2dpdGh1Yi5jb20vaHVnZ2luZ2ZhY2Uva2VybmVscy1jb21tdW5pdHkvLmdpdGh1Yi93b3JrZmxvd3MvYnVpbGQueWFtbEByZWZzL2hlYWRzL21haW4wOQYKKwYBBAGDvzABAQQraHR0cHM6Ly90b2tlbi5hY3Rpb25zLmdpdGh1YnVzZXJjb250ZW50LmNvbTAfBgorBgEEAYO/MAECBBF3b3JrZmxvd19kaXNwYXRjaDA2BgorBgEEAYO/MAEDBCg1ZmU5ZGJiM2Q0NWY2ZjI2NTI5NThjMWY3MGU1MzgxODkxZmE3YWEwMBMGCisGAQQBg78wAQQEBUJ1aWxkMCsGCisGAQQBg78wAQUEHWh1Z2dpbmdmYWNlL2tlcm5lbHMtY29tbXVuaXR5MB0GCisGAQQBg78wAQYED3JlZnMvaGVhZHMvbWFpbjA7BgorBgEEAYO/MAEIBC0MK2h0dHBzOi8vdG9rZW4uYWN0aW9ucy5naXRodWJ1c2VyY29udGVudC5jb20wbQYKKwYBBAGDvzABCQRfDF1odHRwczovL2dpdGh1Yi5jb20vaHVnZ2luZ2ZhY2Uva2VybmVscy1jb21tdW5pdHkvLmdpdGh1Yi93b3JrZmxvd3MvYnVpbGQueWFtbEByZWZzL2hlYWRzL21haW4wOAYKKwYBBAGDvzABCgQqDCg1ZmU5ZGJiM2Q0NWY2ZjI2NTI5NThjMWY3MGU1MzgxODkxZmE3YWEwMBsGCisGAQQBg78wAQsEDQwLc2VsZi1ob3N0ZWQwQAYKKwYBBAGDvzABDAQyDDBodHRwczovL2dpdGh1Yi5jb20vaHVnZ2luZ2ZhY2Uva2VybmVscy1jb21tdW5pdHkwOAYKKwYBBAGDvzABDQQqDCg1ZmU5ZGJiM2Q0NWY2ZjI2NTI5NThjMWY3MGU1MzgxODkxZmE3YWEwMB8GCisGAQQBg78wAQ4EEQwPcmVmcy9oZWFkcy9tYWluMBoGCisGAQQBg78wAQ8EDAwKMTA3MTQ3NTUyOTAuBgorBgEEAYO/MAEQBCAMHmh0dHBzOi8vZ2l0aHViLmNvbS9odWdnaW5nZmFjZTAYBgorBgEEAYO/MAERBAoMCDI1NzIwNzQzMG0GCisGAQQBg78wARIEXwxdaHR0cHM6Ly9naXRodWIuY29tL2h1Z2dpbmdmYWNlL2tlcm5lbHMtY29tbXVuaXR5Ly5naXRodWIvd29ya2Zsb3dzL2J1aWxkLnlhbWxAcmVmcy9oZWFkcy9tYWluMDgGCisGAQQBg78wARMEKgwoNWZlOWRiYjNkNDVmNmYyNjUyOTU4YzFmNzBlNTM4MTg5MWZhN2FhMDAhBgorBgEEAYO/MAEUBBMMEXdvcmtmbG93X2Rpc3BhdGNoMGQGCisGAQQBg78wARUEVgxUaHR0cHM6Ly9naXRodWIuY29tL2h1Z2dpbmdmYWNlL2tlcm5lbHMtY29tbXVuaXR5L2FjdGlvbnMvcnVucy80MTA4MzUwNDUyMC9hdHRlbXB0cy8xMBYGCisGAQQBg78wARYECAwGcHVibGljMEYGCisGAQQBg78wARgEOAw2cmVwbzpodWdnaW5nZmFjZS9rZXJuZWxzLWNvbW11bml0eTpyZWY6cmVmcy9oZWFkcy9tYWluMIGJBgorBgEEAdZ5AgQCBHsEeQB3AHUA3T0wasbHETJjGR4cmWc3AqJKXrjePK3/h4pygC8p7o4AAAGhjn8/ZQAABAMARjBEAiBaAtxdMjhkq4nSffG/MP9zDi0O+2iBzmiZzMQA3+dcHQIgalGrmYXC0ualk7jwFgzDgDg/8qHVaoRfT6pQNKpeeQ0wCgYIKoZIzj0EAwMDaQAwZgIxAJlCmWEsj8r1xO8gB0HTWYwMxOwfeJO2WUBt01LAIq7VXHnQVwci4RmQkjUnJRiaaAIxAL0PbdbDtxvOtvbHPu1SPjcZ92jWDxPVjxwsVXdd5KX9VAC4QTqnhJWAagISaC56ew=="}, "tlogEntries":[{"logIndex":"3601919788", "logId":{"keyId":"wNI9atQGlz+VWfO6LRygH4QUfY/8W4RFwiT5i5WRgB0="}, "kindVersion":{"kind":"hashedrekord", "version":"0.0.1"}, "integratedTime":"1793392068", "inclusionPromise":{"signedEntryTimestamp":"MEUCICslKsOTEg1UIojFLLSfDsOKMES6pqJ14uDZ5NMQmmDkAiEAo0b1yv2n8qaegNOR5cY3CqgrWUcdQrLnVWEY8ggrBNo="}, "inclusionProof":{"logIndex":"3480015526", "rootHash":"e16jcgV5RwnDFrCsepqUJlBde0W+EG9pyUsmuBOqLoY=", "treeSize":"3480017583", "hashes":["kGKvdKdMLjCRdHzpz34FKCF1L5PvT1zRgd8PcN7Am50=", "ku1zmpk/DBUZbQPp7KF+V76rpHjA62SZxiuGCsKswr0=", "0SpeUH6kuW1xfO83OXgGBvBMdWvXNEG4E9QUrBX5i+Y=", "KSqt+4zdICi5lqnMQizRBkBT6+xUWboQqPx/sKcGnio=", "VGCnZlZWMXE/f6nMPfBQNp4s4hgyYs+cSl7bbBD5XuM=", "wyUv+WlVuvo5DZOS/9bfdNdVr1rITY4pRc4xnFNsdDg=", "Bc2r3LfK/zcbdcnofr5I7IVWj/3ELOR32iGT50snxo0=", "EYx4aq3WZ8/X6uVw+UhmbQ6bZ+5jN6Jw5kGt0e8UC5I=", "J8jSls4hipgIB6J393NdOkWAHVHeOp81KR1kyAG7OQo=", "XpQfFCimu9H3vVqGwZsqR5uocsCiUB2Eg2oNRjkK/dw=", "2MrD9kF4W+TvxQ7W34MIbwvuZti6h7Egy0uhRmUPt44=", "GkEJomaVmSyE6uQ2azYoGPn/idtkw3Y0VEXKOrLwfk4=", "N1BSfcRYyYL1rQWHsI3PHVz6aZOqWfiuLsuUuHLC/lM=", "FGzETXd85KSZDiS854wzNAsJ+POTz2es7GVf9cwUo+s=", "oRoie0t/H+0ZyKRjI6KiDz9stRDDR7EJu7t8qrlVT0Y=", "a8cmhOLBStNzSIVybWzlBd1UxGct2Vy2Eggt5g2B26w=", "39Z7bKKSjPtMpTU8xt8EybgTkCg1qk72XWkQqE5WRSU=", "63Z7Fze4Sm0atCwhCb6EG/N+w2qm+Q+7+dWxf2+uPzA=", "njdy0pCf5uBXEpRPrJGvuVRc6ibqank2dtzoCQ0SR18=", "e3OaWwzL1tEWYXiUzGpH+1v9CiFPwweyfRzIh3Z+uO4=", "b+938HIdfjdD/+LgIs3AVRE3Wc65/1NSSAL4OJ5FUXE=", "kQz/4ND7LdhfAQn+idNsO+ZFC32sfkYRlhjWF5W+sPc=", "zGfeoWKmrZlXMOvyi3R2GDbGpmI1skJW/wLTXCCgF0o=", "1DaDKzrCNNBhmSdY5vFjz+NpAmxmYfYEgBCdPLrMP1A=", "xH/DCseLHr9eKoYT8qsORZK7zVdEGYWHuVtsVrD95wY="], "checkpoint":{"envelope":"rekor.sigstore.dev - 1193050959916656506\n3480017583\ne16jcgV5RwnDFrCsepqUJlBde0W+EG9pyUsmuBOqLoY=\n\n— rekor.sigstore.dev wNI9ajBEAiA2/PEKD7vlrcIDXEOALf8L61IhKh+aBibat8qHWySxSAIgFshPX03SFKXqpKlEvasGC0NxlndNMe/pZvUbc/C3Kvs=\n"}}, "canonicalizedBody":"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"}], "timestampVerificationData":{"rfc3161Timestamps":[{"signedTimestamp":"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"}]}}, "messageSignature":{"messageDigest":{"algorithm":"SHA2_256", "digest":"codCV4kUGio3T6/NE1If9ZY5lT7h8AU2aBRUMl0HuYk="}, "signature":"MEUCICfQIrklgvKvmi3QgHAkLUNfbeb0H31uwWTPTeCE9C/YAiEAnCkgBKN9PzDy2BsQcfwbNnQgv1zRvKhSosmDQ9Az78c="}}
|
|
|
|
| 1 |
+
{"mediaType":"application/vnd.dev.sigstore.bundle.v0.3+json", "verificationMaterial":{"certificate":{"rawBytes":"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"}, "tlogEntries":[{"logIndex":"3654776552", "logId":{"keyId":"wNI9atQGlz+VWfO6LRygH4QUfY/8W4RFwiT5i5WRgB0="}, "kindVersion":{"kind":"hashedrekord", "version":"0.0.1"}, "integratedTime":"1793669010", "inclusionPromise":{"signedEntryTimestamp":"MEQCIDi+fxpQfAZL6BXxHdgUxSM37e90HD49f0J1hsspi8AJAiBS6c0tPmUbchcsdSZNnS6szTTXaxZtpBMfGoM3LoJX2g=="}, "inclusionProof":{"logIndex":"3532872290", "rootHash":"K6znMRF/3WkE9HsYssR7RbZ/4uaYbPgbliGau903UM0=", "treeSize":"3532872456", "hashes":["b/7JcHGBy2J+i9IvxcMJn80o0eqyVTcDGOUfyevnoWU=", "WThaaVGKZXscicuutjVM47+LYUpKslc+HS/KCSDuOyo=", "9TgQCSaWStL+BDex4HdZyGOuXGowIxpZ7bI3Vn8ezqo=", "UXtA9+YxSsZKrC3t6Rvjbc9x5IQmUFWFszzCSejSiic=", "wYo3/28RXZ4gwyjGFHgbzwmgxaVYsKAXugADq5x1ltg=", "QXoeLTjpEeA9ROS8JBlWLenMvOScBZbta2og6zeFMNQ=", "auWTv4l34KoVYBtd22J3tKoMaDVwPPxwlduBzPLN3xU=", "S9wcLaJK5L1iEZegbK1YvOmfWbf6sRd9d6akhpkAQhc=", "nSiFIL7QAyp2/8v/bp5jD4uzJ+s0xAspIUIuATjLGAs=", "iQIroKz45eghtOUtBTkEKvrA5P/YwR3rC4es/hpQYlQ=", "FDXHnU81UyrxLLXECQOkKxh59vRdnnsos1VpjA1XfsY=", "l2F8XDZWSpFsRQNOEJbYK/sHi0FHC7rqAIaBXYhU1pQ=", "jXRnHZhdIEDN/dOSOGnahiPHiCL8svXXZkNlpklSSDE=", "JgdHnyIJAAm3PA7mCUWsdLQ/ZPPYPC6Oh1Gw79WHtpg=", "MlqLoiFm8Oo5bbmftEQbgcSKXfa+i5FCiungGzOC1uo=", "QYcxedNInCRxBCNHkv8RT2T790mUNX70dZsQ78gLIYQ=", "ogjRZ88pnEW/ANOWeLXfhYL/XdgeGJy0zJTgCayJlGE=", "PA9MhoiExYPp/vXh8tVCBrATkoY31KStgze+rovDTkA=", "SlixnS34BhAO+yRI2FaGPSOh9V6JSVNPfLl1bHxrcQY=", "1DaDKzrCNNBhmSdY5vFjz+NpAmxmYfYEgBCdPLrMP1A=", "xH/DCseLHr9eKoYT8qsORZK7zVdEGYWHuVtsVrD95wY="], "checkpoint":{"envelope":"rekor.sigstore.dev - 1193050959916656506\n3532872456\nK6znMRF/3WkE9HsYssR7RbZ/4uaYbPgbliGau903UM0=\n\n— rekor.sigstore.dev wNI9ajBFAiEAzbgwg9jpKhjvILMOoaT3sOOChiLWhKf6+Se/wC8BA5cCID7SauW1U7RUVqKPJl1Rhncq96J4HBxGY9rafX7f43xT\n"}}, "canonicalizedBody":"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"}], "timestampVerificationData":{"rfc3161Timestamps":[{"signedTimestamp":"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"}]}}, "messageSignature":{"messageDigest":{"algorithm":"SHA2_256", "digest":"AfilE90u7mtztHU6TfI+rIb0pDjSkF7q/7d04r6UjeE="}, "signature":"MEQCID853i3gHmRIO6Ydr5d+Gg9L88A3sgo6bNdrEuCJNqCiAiBzErOcF5DyQteZDCbVYKQGRZ0ciLETeGdl3UsAp8xiAQ=="}}
|
build/torch-stable-abi29-cu128-x86_64-linux/symbols.json
CHANGED
|
@@ -1,11 +1,11 @@
|
|
| 1 |
{
|
| 2 |
"schema_version": 1,
|
| 3 |
-
"module": "
|
| 4 |
"functions": [
|
| 5 |
{
|
| 6 |
"name": "add_op_namespace_prefix",
|
| 7 |
"qualname": "add_op_namespace_prefix",
|
| 8 |
-
"module": "
|
| 9 |
"kind": "function",
|
| 10 |
"doc": "Prefix op by namespace.",
|
| 11 |
"signature": {
|
|
@@ -23,7 +23,7 @@
|
|
| 23 |
{
|
| 24 |
"name": "flash_attn_combine",
|
| 25 |
"qualname": "flash_attn_combine",
|
| 26 |
-
"module": "
|
| 27 |
"kind": "function",
|
| 28 |
"doc": null,
|
| 29 |
"signature": {
|
|
@@ -59,7 +59,7 @@
|
|
| 59 |
{
|
| 60 |
"name": "flash_attn_func",
|
| 61 |
"qualname": "flash_attn_func",
|
| 62 |
-
"module": "
|
| 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": {
|
|
@@ -173,7 +173,7 @@
|
|
| 173 |
{
|
| 174 |
"name": "flash_attn_qkvpacked_func",
|
| 175 |
"qualname": "flash_attn_qkvpacked_func",
|
| 176 |
-
"module": "
|
| 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": {
|
|
@@ -263,7 +263,7 @@
|
|
| 263 |
{
|
| 264 |
"name": "flash_attn_varlen_func",
|
| 265 |
"qualname": "flash_attn_varlen_func",
|
| 266 |
-
"module": "
|
| 267 |
"kind": "function",
|
| 268 |
"doc": null,
|
| 269 |
"signature": {
|
|
@@ -413,7 +413,7 @@
|
|
| 413 |
{
|
| 414 |
"name": "flash_attn_with_kvcache",
|
| 415 |
"qualname": "flash_attn_with_kvcache",
|
| 416 |
-
"module": "
|
| 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": {
|
|
@@ -605,7 +605,7 @@
|
|
| 605 |
{
|
| 606 |
"name": "get_scheduler_metadata",
|
| 607 |
"qualname": "get_scheduler_metadata",
|
| 608 |
-
"module": "
|
| 609 |
"kind": "function",
|
| 610 |
"doc": null,
|
| 611 |
"signature": {
|
|
|
|
| 1 |
{
|
| 2 |
"schema_version": 1,
|
| 3 |
+
"module": "_flash_attn3_cuda_1ac140b",
|
| 4 |
"functions": [
|
| 5 |
{
|
| 6 |
"name": "add_op_namespace_prefix",
|
| 7 |
"qualname": "add_op_namespace_prefix",
|
| 8 |
+
"module": "_flash_attn3_cuda_1ac140b._ops",
|
| 9 |
"kind": "function",
|
| 10 |
"doc": "Prefix op by namespace.",
|
| 11 |
"signature": {
|
|
|
|
| 23 |
{
|
| 24 |
"name": "flash_attn_combine",
|
| 25 |
"qualname": "flash_attn_combine",
|
| 26 |
+
"module": "_flash_attn3_cuda_1ac140b.flash_attn_interface",
|
| 27 |
"kind": "function",
|
| 28 |
"doc": null,
|
| 29 |
"signature": {
|
|
|
|
| 59 |
{
|
| 60 |
"name": "flash_attn_func",
|
| 61 |
"qualname": "flash_attn_func",
|
| 62 |
+
"module": "_flash_attn3_cuda_1ac140b.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": {
|
|
|
|
| 173 |
{
|
| 174 |
"name": "flash_attn_qkvpacked_func",
|
| 175 |
"qualname": "flash_attn_qkvpacked_func",
|
| 176 |
+
"module": "_flash_attn3_cuda_1ac140b.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": {
|
|
|
|
| 263 |
{
|
| 264 |
"name": "flash_attn_varlen_func",
|
| 265 |
"qualname": "flash_attn_varlen_func",
|
| 266 |
+
"module": "_flash_attn3_cuda_1ac140b.flash_attn_interface",
|
| 267 |
"kind": "function",
|
| 268 |
"doc": null,
|
| 269 |
"signature": {
|
|
|
|
| 413 |
{
|
| 414 |
"name": "flash_attn_with_kvcache",
|
| 415 |
"qualname": "flash_attn_with_kvcache",
|
| 416 |
+
"module": "_flash_attn3_cuda_1ac140b.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": {
|
|
|
|
| 605 |
{
|
| 606 |
"name": "get_scheduler_metadata",
|
| 607 |
"qualname": "get_scheduler_metadata",
|
| 608 |
+
"module": "_flash_attn3_cuda_1ac140b.flash_attn_interface",
|
| 609 |
"kind": "function",
|
| 610 |
"doc": null,
|
| 611 |
"signature": {
|
build/torch-stable-abi29-cu130-x86_64-linux/{_flash_attn3_cuda_5fe9dbb.abi3.so → _flash_attn3_cuda_1ac140b.abi3.so}
RENAMED
|
@@ -1,3 +1,3 @@
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
-
oid sha256:
|
| 3 |
size 549647192
|
|
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:fab9715dab1fef23fc46ebc993e01f03c32d4a839a023cb2c870abd3d2abe479
|
| 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_1ac140b
|
| 3 |
+
ops = torch.ops._flash_attn3_cuda_1ac140b
|
| 4 |
|
| 5 |
def add_op_namespace_prefix(op_name: str):
|
| 6 |
"""
|
| 7 |
Prefix op by namespace.
|
| 8 |
"""
|
| 9 |
+
return f"_flash_attn3_cuda_1ac140b::{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",
|
|
@@ -17,21 +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 |
-
"symbols.json": "
|
| 25 |
}
|
| 26 |
},
|
| 27 |
"provenance": {
|
| 28 |
"kernel-builder": {
|
| 29 |
-
"version": "0.
|
| 30 |
-
"commit": "
|
| 31 |
"dirty": false
|
| 32 |
},
|
| 33 |
"kernel": {
|
| 34 |
-
"commit": "
|
| 35 |
"dirty": false
|
| 36 |
}
|
| 37 |
}
|
|
|
|
| 1 |
{
|
| 2 |
"name": "flash-attn3",
|
| 3 |
+
"id": "_flash_attn3_cuda_1ac140b",
|
| 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_1ac140b.abi3.so": "+rlxXasf7yP8RuvJk+AfA8MtSoOaAjyyyHCr09Kr5Hk=",
|
| 21 |
+
"_ops.py": "NH6/V6V8u1HinwVJvVwoG30z34Aa9OL/sgYZocQi7a8=",
|
| 22 |
"flash_attn_config.py": "uxo+eyMcDit//8YDaS4Rtk5BTMKzTSK9tOKTjdWVe/w=",
|
| 23 |
"flash_attn_interface.py": "y0bD2JYGMFgX6iDre3zJN2sBPuq/12N6Mo55Z+HQCNQ=",
|
| 24 |
+
"symbols.json": "qzII4ieoZRHwN7dHbNFGa995mlTpWyJs1X5eRdNLU5o="
|
| 25 |
}
|
| 26 |
},
|
| 27 |
"provenance": {
|
| 28 |
"kernel-builder": {
|
| 29 |
+
"version": "0.19.0-dev0",
|
| 30 |
+
"commit": "88e9dd07eb4e639b85799c14b2d8af14a0d9580a",
|
| 31 |
"dirty": false
|
| 32 |
},
|
| 33 |
"kernel": {
|
| 34 |
+
"commit": "1ac140bf36af3c8ff0e73262add55fb8df939e26",
|
| 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":"MIIHKzCCBrGgAwIBAgIUfkY4VMpmE83aJLdGgE9BIs96E3IwCgYIKoZIzj0EAwMwNzEVMBMGA1UEChMMc2lnc3RvcmUuZGV2MR4wHAYDVQQDExVzaWdzdG9yZS1pbnRlcm1lZGlhdGUwHhcNMjYxMDMwMjAyNzI4WhcNMjYxMDMwMjAzNzI4WjAAMFkwEwYHKoZIzj0CAQYIKoZIzj0DAQcDQgAEi0C4IYE7FSNOxOfjXJJzKSqWk4DVOeQIHiCeoltlcG6gfnSwzMjfSbwMy1Ya7mTI0ouqwexDD4GITFyfg+xlR6OCBdAwggXMMA4GA1UdDwEB/wQEAwIHgDATBgNVHSUEDDAKBggrBgEFBQcDAzAfBgNVHSMEGDAWgBTf0+nPViQRlvmo2OkoVaLGLhhkPzBrBgNVHREBAf8EYTBfhl1odHRwczovL2dpdGh1Yi5jb20vaHVnZ2luZ2ZhY2Uva2VybmVscy1jb21tdW5pdHkvLmdpdGh1Yi93b3JrZmxvd3MvYnVpbGQueWFtbEByZWZzL2hlYWRzL21haW4wOQYKKwYBBAGDvzABAQQraHR0cHM6Ly90b2tlbi5hY3Rpb25zLmdpdGh1YnVzZXJjb250ZW50LmNvbTAfBgorBgEEAYO/MAECBBF3b3JrZmxvd19kaXNwYXRjaDA2BgorBgEEAYO/MAEDBCg1ZmU5ZGJiM2Q0NWY2ZjI2NTI5NThjMWY3MGU1MzgxODkxZmE3YWEwMBMGCisGAQQBg78wAQQEBUJ1aWxkMCsGCisGAQQBg78wAQUEHWh1Z2dpbmdmYWNlL2tlcm5lbHMtY29tbXVuaXR5MB0GCisGAQQBg78wAQYED3JlZnMvaGVhZHMvbWFpbjA7BgorBgEEAYO/MAEIBC0MK2h0dHBzOi8vdG9rZW4uYWN0aW9ucy5naXRodWJ1c2VyY29udGVudC5jb20wbQYKKwYBBAGDvzABCQRfDF1odHRwczovL2dpdGh1Yi5jb20vaHVnZ2luZ2ZhY2Uva2VybmVscy1jb21tdW5pdHkvLmdpdGh1Yi93b3JrZmxvd3MvYnVpbGQueWFtbEByZWZzL2hlYWRzL21haW4wOAYKKwYBBAGDvzABCgQqDCg1ZmU5ZGJiM2Q0NWY2ZjI2NTI5NThjMWY3MGU1MzgxODkxZmE3YWEwMBsGCisGAQQBg78wAQsEDQwLc2VsZi1ob3N0ZWQwQAYKKwYBBAGDvzABDAQyDDBodHRwczovL2dpdGh1Yi5jb20vaHVnZ2luZ2ZhY2Uva2VybmVscy1jb21tdW5pdHkwOAYKKwYBBAGDvzABDQQqDCg1ZmU5ZGJiM2Q0NWY2ZjI2NTI5NThjMWY3MGU1MzgxODkxZmE3YWEwMB8GCisGAQQBg78wAQ4EEQwPcmVmcy9oZWFkcy9tYWluMBoGCisGAQQBg78wAQ8EDAwKMTA3MTQ3NTUyOTAuBgorBgEEAYO/MAEQBCAMHmh0dHBzOi8vZ2l0aHViLmNvbS9odWdnaW5nZmFjZTAYBgorBgEEAYO/MAERBAoMCDI1NzIwNzQzMG0GCisGAQQBg78wARIEXwxdaHR0cHM6Ly9naXRodWIuY29tL2h1Z2dpbmdmYWNlL2tlcm5lbHMtY29tbXVuaXR5Ly5naXRodWIvd29ya2Zsb3dzL2J1aWxkLnlhbWxAcmVmcy9oZWFkcy9tYWluMDgGCisGAQQBg78wARMEKgwoNWZlOWRiYjNkNDVmNmYyNjUyOTU4YzFmNzBlNTM4MTg5MWZhN2FhMDAhBgorBgEEAYO/MAEUBBMMEXdvcmtmbG93X2Rpc3BhdGNoMGQGCisGAQQBg78wARUEVgxUaHR0cHM6Ly9naXRodWIuY29tL2h1Z2dpbmdmYWNlL2tlcm5lbHMtY29tbXVuaXR5L2FjdGlvbnMvcnVucy80MTA4MzUwNDUyMC9hdHRlbXB0cy8xMBYGCisGAQQBg78wARYECAwGcHVibGljMEYGCisGAQQBg78wARgEOAw2cmVwbzpodWdnaW5nZmFjZS9rZXJuZWxzLWNvbW11bml0eTpyZWY6cmVmcy9oZWFkcy9tYWluMIGJBgorBgEEAdZ5AgQCBHsEeQB3AHUA3T0wasbHETJjGR4cmWc3AqJKXrjePK3/h4pygC8p7o4AAAGhjn748AAABAMARjBEAiAKCRMt2VTlkqpXe2SI/6DkyZbF3esBNcQK3PJaAfcEhwIgZgWhRDhYmeBhIMySoR4L0EMoEdRZPjSO0Xyzbc461K0wCgYIKoZIzj0EAwMDaAAwZQIwR2zgke3wlCFEm9mDN7UlhlMsrlC1wMgscG01DqHboql1wnj2gqGgTkh7eF7DpukXAjEAlVeSAFW4pjg2BEL3kgAYRfx25+AfhMWjgEGh2Dd2/EWgP0FP2gl6dv9FveZS1WAI"}, "tlogEntries":[{"logIndex":"3601913433", "logId":{"keyId":"wNI9atQGlz+VWfO6LRygH4QUfY/8W4RFwiT5i5WRgB0="}, "kindVersion":{"kind":"hashedrekord", "version":"0.0.1"}, "integratedTime":"1793392050", "inclusionPromise":{"signedEntryTimestamp":"MEUCIG6iec8BOBxkCJgFgA6cYYFeFaT2WeRgQ91Gs4lDYeX4AiEA9eW4F+5t4YdFSgFNqt8gjChFLcgrUNRk4SjP1hxuha0="}, "inclusionProof":{"logIndex":"3480009171", "rootHash":"/U7cq5ZYhY9L3LBglkXwFcv4+JlhYxR2tgDr+eFOOxc=", "treeSize":"3480009629", "hashes":["E7JBFyXJx9QhhPn/wBHI3ppTYDAPowkrLUkSiRQ/duw=", "6HNHVe6vYK9zlW6N8v/a65q0dzJ9OcBsXYAhBrf3QVI=", "iLZfQgDMR22bGAHe9z+fYycJBZ5VltGNSHI7axPrInQ=", "STY7rjtl1RUElmtd2xYnKa9ADi9cKleDm9+6wWqI4XQ=", "p4lZ0/Jw8dxsRgX0RIzrea6zT/SfxoY+eVCwi0mv+PA=", "dcjAFHsbF57PbFva08cMdEu6GY6j1TLhQXwioCc0QLo=", "iwyWNDCTesEXH69lKy0C3hrTi1mTQ9NL7PLDGG4U3zg=", "TSN6InB6/wNFtZczBzVr3JWx3BRwyf9oK+HyEOvq9Qs=", "tMvRBQwYjdw/s7CBtyNGHWHAVKVIxup4+Igv0P2mZOA=", "3aTRjGbHA5Hlti6M7PV56Amxwarwr3I9/aHhrFwKGLk=", "xij9vw5U2ZmpBO0ooE/cTzfl0A8+1ifweHjmiSMLWr8=", "mUtKEp1REU8ipx4txkRb2EFZ4aOx/bgS3JxRL01oG54=", "9t3kgP+u+MFv895vMJiyVlUx0Z9SHSeUhxjsfdwGIos=", "oRoie0t/H+0ZyKRjI6KiDz9stRDDR7EJu7t8qrlVT0Y=", "a8cmhOLBStNzSIVybWzlBd1UxGct2Vy2Eggt5g2B26w=", "39Z7bKKSjPtMpTU8xt8EybgTkCg1qk72XWkQqE5WRSU=", "63Z7Fze4Sm0atCwhCb6EG/N+w2qm+Q+7+dWxf2+uPzA=", "njdy0pCf5uBXEpRPrJGvuVRc6ibqank2dtzoCQ0SR18=", "e3OaWwzL1tEWYXiUzGpH+1v9CiFPwweyfRzIh3Z+uO4=", "b+938HIdfjdD/+LgIs3AVRE3Wc65/1NSSAL4OJ5FUXE=", "kQz/4ND7LdhfAQn+idNsO+ZFC32sfkYRlhjWF5W+sPc=", "zGfeoWKmrZlXMOvyi3R2GDbGpmI1skJW/wLTXCCgF0o=", "1DaDKzrCNNBhmSdY5vFjz+NpAmxmYfYEgBCdPLrMP1A=", "xH/DCseLHr9eKoYT8qsORZK7zVdEGYWHuVtsVrD95wY="], "checkpoint":{"envelope":"rekor.sigstore.dev - 1193050959916656506\n3480009629\n/U7cq5ZYhY9L3LBglkXwFcv4+JlhYxR2tgDr+eFOOxc=\n\n— rekor.sigstore.dev wNI9ajBGAiEAsRmDPqOn0RRnnlYSuLBWh/92MldZMj6V9hV4QT8H1rUCIQDP4lEO1JnYtCrRVEpYu0lT5j+Hy2lJHpLxl5Rt9xNePg==\n"}}, "canonicalizedBody":"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"}], "timestampVerificationData":{"rfc3161Timestamps":[{"signedTimestamp":"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"}]}}, "messageSignature":{"messageDigest":{"algorithm":"SHA2_256", "digest":"hh3iBC0Lurii2gA7FA+ylxJcV4taQW1BQECnzF3AEEo="}, "signature":"MEQCIGeiiYeRIHWSZfV8YAfHvKDqMX1uBTK6uYbvcHqKl0kQAiAzrBHSHQlKn/TLwrXcPLT/okbNjgMY6Tz+ngt5IbnB+g=="}}
|
|
|
|
| 1 |
+
{"mediaType":"application/vnd.dev.sigstore.bundle.v0.3+json", "verificationMaterial":{"certificate":{"rawBytes":"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"}, "tlogEntries":[{"logIndex":"3654777340", "logId":{"keyId":"wNI9atQGlz+VWfO6LRygH4QUfY/8W4RFwiT5i5WRgB0="}, "kindVersion":{"kind":"hashedrekord", "version":"0.0.1"}, "integratedTime":"1793669013", "inclusionPromise":{"signedEntryTimestamp":"MEQCIGAPD9yTK19z+HUX0AeE5xTfjUP/ooMWWWS5d5CwuN4iAiB9PXOWiwcEybDfJikIPRVjFultlRFvtltTpRBP787Blg=="}, "inclusionProof":{"logIndex":"3532873078", "rootHash":"Ap/CT2B1yuIWaenPEH1bE5PKDuYhe3hEO/gxYA+wqQ4=", "treeSize":"3532873148", "hashes":["Adv9m/KswTsOUgHP380LsIopccxhHOIFkls1vM9CtDo=", "piwcOakSux+LWfQRLY/J2YtkBmGQNF7yJH7s1LGIFZI=", "HQccA7ql9uipOuw7YxTfJzzzGGJ12gSE4tu5D97C7g0=", "ASTB7yR4TLbTVK9bjHPYyNRdp8XHYiHu8iDdNno08NE=", "X0n/gK4695/+GWLALe2ubL4Y2hjZGhpTTuWtQ8i/F+8=", "Hj00qS+IVrMyxR61lovjUBBKrkr3RJsvQnOAcHvT2ak=", "LQwKC0NdoJYnghRtdooitfr9iVJKLzrtIIrkP+zodxg=", "5nsNBdDFP+nigBSxmtjEqXb3kPZ6YNF9PgS4Llo7lQM=", "TFb04ogJlrWvpwINzYTIOnfR+WSeSQKne6lcYJ7vlGs=", "bIGmaTiU2b0sOuw7AvrHIPr5/wzqwwFxz3MVAKSq5Vo=", "FDXHnU81UyrxLLXECQOkKxh59vRdnnsos1VpjA1XfsY=", "l2F8XDZWSpFsRQNOEJbYK/sHi0FHC7rqAIaBXYhU1pQ=", "jXRnHZhdIEDN/dOSOGnahiPHiCL8svXXZkNlpklSSDE=", "JgdHnyIJAAm3PA7mCUWsdLQ/ZPPYPC6Oh1Gw79WHtpg=", "MlqLoiFm8Oo5bbmftEQbgcSKXfa+i5FCiungGzOC1uo=", "QYcxedNInCRxBCNHkv8RT2T790mUNX70dZsQ78gLIYQ=", "ogjRZ88pnEW/ANOWeLXfhYL/XdgeGJy0zJTgCayJlGE=", "PA9MhoiExYPp/vXh8tVCBrATkoY31KStgze+rovDTkA=", "SlixnS34BhAO+yRI2FaGPSOh9V6JSVNPfLl1bHxrcQY=", "1DaDKzrCNNBhmSdY5vFjz+NpAmxmYfYEgBCdPLrMP1A=", "xH/DCseLHr9eKoYT8qsORZK7zVdEGYWHuVtsVrD95wY="], "checkpoint":{"envelope":"rekor.sigstore.dev - 1193050959916656506\n3532873148\nAp/CT2B1yuIWaenPEH1bE5PKDuYhe3hEO/gxYA+wqQ4=\n\n— rekor.sigstore.dev wNI9ajBFAiEA72j0fYlz/SLc9lm6t5D3evrzh0CErFF+U9pRU8ctD7ICIHhp8jgFFSDsfzvJSuers7Hlh6zR5lBls0h3j27K4EbV\n"}}, "canonicalizedBody":"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"}], "timestampVerificationData":{"rfc3161Timestamps":[{"signedTimestamp":"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"}]}}, "messageSignature":{"messageDigest":{"algorithm":"SHA2_256", "digest":"X+6q6yHSdIBFitX/tFaa9pWFvPgb5H6XNzxkC463Hp0="}, "signature":"MEYCIQDv/yC4iQRWvOzz7hy19QPlTCLvna2oDwwEIyfulRMtMgIhANXv1YXvwVQ0t0OcfSEPWQVpv4cUSntXhGPDxhT0dZlR"}}
|
build/torch-stable-abi29-cu130-x86_64-linux/symbols.json
CHANGED
|
@@ -1,11 +1,11 @@
|
|
| 1 |
{
|
| 2 |
"schema_version": 1,
|
| 3 |
-
"module": "
|
| 4 |
"functions": [
|
| 5 |
{
|
| 6 |
"name": "add_op_namespace_prefix",
|
| 7 |
"qualname": "add_op_namespace_prefix",
|
| 8 |
-
"module": "
|
| 9 |
"kind": "function",
|
| 10 |
"doc": "Prefix op by namespace.",
|
| 11 |
"signature": {
|
|
@@ -23,7 +23,7 @@
|
|
| 23 |
{
|
| 24 |
"name": "flash_attn_combine",
|
| 25 |
"qualname": "flash_attn_combine",
|
| 26 |
-
"module": "
|
| 27 |
"kind": "function",
|
| 28 |
"doc": null,
|
| 29 |
"signature": {
|
|
@@ -59,7 +59,7 @@
|
|
| 59 |
{
|
| 60 |
"name": "flash_attn_func",
|
| 61 |
"qualname": "flash_attn_func",
|
| 62 |
-
"module": "
|
| 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": {
|
|
@@ -173,7 +173,7 @@
|
|
| 173 |
{
|
| 174 |
"name": "flash_attn_qkvpacked_func",
|
| 175 |
"qualname": "flash_attn_qkvpacked_func",
|
| 176 |
-
"module": "
|
| 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": {
|
|
@@ -263,7 +263,7 @@
|
|
| 263 |
{
|
| 264 |
"name": "flash_attn_varlen_func",
|
| 265 |
"qualname": "flash_attn_varlen_func",
|
| 266 |
-
"module": "
|
| 267 |
"kind": "function",
|
| 268 |
"doc": null,
|
| 269 |
"signature": {
|
|
@@ -413,7 +413,7 @@
|
|
| 413 |
{
|
| 414 |
"name": "flash_attn_with_kvcache",
|
| 415 |
"qualname": "flash_attn_with_kvcache",
|
| 416 |
-
"module": "
|
| 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": {
|
|
@@ -605,7 +605,7 @@
|
|
| 605 |
{
|
| 606 |
"name": "get_scheduler_metadata",
|
| 607 |
"qualname": "get_scheduler_metadata",
|
| 608 |
-
"module": "
|
| 609 |
"kind": "function",
|
| 610 |
"doc": null,
|
| 611 |
"signature": {
|
|
|
|
| 1 |
{
|
| 2 |
"schema_version": 1,
|
| 3 |
+
"module": "_flash_attn3_cuda_1ac140b",
|
| 4 |
"functions": [
|
| 5 |
{
|
| 6 |
"name": "add_op_namespace_prefix",
|
| 7 |
"qualname": "add_op_namespace_prefix",
|
| 8 |
+
"module": "_flash_attn3_cuda_1ac140b._ops",
|
| 9 |
"kind": "function",
|
| 10 |
"doc": "Prefix op by namespace.",
|
| 11 |
"signature": {
|
|
|
|
| 23 |
{
|
| 24 |
"name": "flash_attn_combine",
|
| 25 |
"qualname": "flash_attn_combine",
|
| 26 |
+
"module": "_flash_attn3_cuda_1ac140b.flash_attn_interface",
|
| 27 |
"kind": "function",
|
| 28 |
"doc": null,
|
| 29 |
"signature": {
|
|
|
|
| 59 |
{
|
| 60 |
"name": "flash_attn_func",
|
| 61 |
"qualname": "flash_attn_func",
|
| 62 |
+
"module": "_flash_attn3_cuda_1ac140b.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": {
|
|
|
|
| 173 |
{
|
| 174 |
"name": "flash_attn_qkvpacked_func",
|
| 175 |
"qualname": "flash_attn_qkvpacked_func",
|
| 176 |
+
"module": "_flash_attn3_cuda_1ac140b.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": {
|
|
|
|
| 263 |
{
|
| 264 |
"name": "flash_attn_varlen_func",
|
| 265 |
"qualname": "flash_attn_varlen_func",
|
| 266 |
+
"module": "_flash_attn3_cuda_1ac140b.flash_attn_interface",
|
| 267 |
"kind": "function",
|
| 268 |
"doc": null,
|
| 269 |
"signature": {
|
|
|
|
| 413 |
{
|
| 414 |
"name": "flash_attn_with_kvcache",
|
| 415 |
"qualname": "flash_attn_with_kvcache",
|
| 416 |
+
"module": "_flash_attn3_cuda_1ac140b.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": {
|
|
|
|
| 605 |
{
|
| 606 |
"name": "get_scheduler_metadata",
|
| 607 |
"qualname": "get_scheduler_metadata",
|
| 608 |
+
"module": "_flash_attn3_cuda_1ac140b.flash_attn_interface",
|
| 609 |
"kind": "function",
|
| 610 |
"doc": null,
|
| 611 |
"signature": {
|
build/torch-stable-abi29-cu132-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-cu132-x86_64-linux/_flash_attn3_cuda_1ac140b.abi3.so
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:133f235518caa87d3bee68ac04085eae2605db2a5339bb23ae3d3bc8e0cdcc72
|
| 3 |
+
size 549366160
|
build/torch-stable-abi29-cu132-x86_64-linux/_ops.py
ADDED
|
@@ -0,0 +1,9 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import torch
|
| 2 |
+
from . import _flash_attn3_cuda_1ac140b
|
| 3 |
+
ops = torch.ops._flash_attn3_cuda_1ac140b
|
| 4 |
+
|
| 5 |
+
def add_op_namespace_prefix(op_name: str):
|
| 6 |
+
"""
|
| 7 |
+
Prefix op by namespace.
|
| 8 |
+
"""
|
| 9 |
+
return f"_flash_attn3_cuda_1ac140b::{op_name}"
|
build/torch-stable-abi29-cu132-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-cu132-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-cu132-x86_64-linux/metadata.json
ADDED
|
@@ -0,0 +1,38 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"name": "flash-attn3",
|
| 3 |
+
"id": "_flash_attn3_cuda_1ac140b",
|
| 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_1ac140b.abi3.so": "Ez8jVRjKqH077misBAheriYF2ypTObsjrj07yODNzHI=",
|
| 21 |
+
"_ops.py": "NH6/V6V8u1HinwVJvVwoG30z34Aa9OL/sgYZocQi7a8=",
|
| 22 |
+
"flash_attn_config.py": "uxo+eyMcDit//8YDaS4Rtk5BTMKzTSK9tOKTjdWVe/w=",
|
| 23 |
+
"flash_attn_interface.py": "y0bD2JYGMFgX6iDre3zJN2sBPuq/12N6Mo55Z+HQCNQ=",
|
| 24 |
+
"symbols.json": "qzII4ieoZRHwN7dHbNFGa995mlTpWyJs1X5eRdNLU5o="
|
| 25 |
+
}
|
| 26 |
+
},
|
| 27 |
+
"provenance": {
|
| 28 |
+
"kernel-builder": {
|
| 29 |
+
"version": "0.19.0-dev0",
|
| 30 |
+
"commit": "88e9dd07eb4e639b85799c14b2d8af14a0d9580a",
|
| 31 |
+
"dirty": false
|
| 32 |
+
},
|
| 33 |
+
"kernel": {
|
| 34 |
+
"commit": "1ac140bf36af3c8ff0e73262add55fb8df939e26",
|
| 35 |
+
"dirty": false
|
| 36 |
+
}
|
| 37 |
+
}
|
| 38 |
+
}
|
build/torch-stable-abi29-cu132-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":"3654778044", "logId":{"keyId":"wNI9atQGlz+VWfO6LRygH4QUfY/8W4RFwiT5i5WRgB0="}, "kindVersion":{"kind":"hashedrekord", "version":"0.0.1"}, "integratedTime":"1793669015", "inclusionPromise":{"signedEntryTimestamp":"MEQCIH1JgPRC8f6RaDxcbVv7G5lLBdx28lxnq2gYA2zu3yOZAiA8Wtm85WzOixUsYGUu9gQr+N688cWDr1YzQK69KkzNCw=="}, "inclusionProof":{"logIndex":"3532873782", "rootHash":"AirPSjrVfD4Np43u5BUpH8irmCFP7voALlh00GvOOek=", "treeSize":"3532873890", "hashes":["+PZYhuz8h1vB48nNWqtoOZFlhMNSv/Xx9DnnQs6KqeU=", "HPkOSXqEek4I6sP/oKxrXwAtHIXCoe5abspbK90c+aE=", "Uyu4AuOcystuW8l+3I6iCWrGjuDZf+4K/BpOMcGJS70=", "ZGPteusGAukYNmRjC02jjRs2UzXjOROvQXoJG+MrjoI=", "Y8udMRDoppTWR6BwUo7sDpLUnzk/VudwxrTfDbYS7jk=", "BUlHrrCcqsDa8dzP7ZAa0K7eSwlzq68O3myyk6ljorI=", "75dHsizhhvyWQqFQkoovc3qkLQzwYwBl6B1S13YlEaM=", "78aTnGQUgpGijhBk09TIHveEghenvNutGckg13ylTww=", "VlrHGUxNk5PYgNCjDyZFwri2UY985FoN1fBJ2z6y/Aw=", "jXRnHZhdIEDN/dOSOGnahiPHiCL8svXXZkNlpklSSDE=", "JgdHnyIJAAm3PA7mCUWsdLQ/ZPPYPC6Oh1Gw79WHtpg=", "MlqLoiFm8Oo5bbmftEQbgcSKXfa+i5FCiungGzOC1uo=", "QYcxedNInCRxBCNHkv8RT2T790mUNX70dZsQ78gLIYQ=", "ogjRZ88pnEW/ANOWeLXfhYL/XdgeGJy0zJTgCayJlGE=", "PA9MhoiExYPp/vXh8tVCBrATkoY31KStgze+rovDTkA=", "SlixnS34BhAO+yRI2FaGPSOh9V6JSVNPfLl1bHxrcQY=", "1DaDKzrCNNBhmSdY5vFjz+NpAmxmYfYEgBCdPLrMP1A=", "xH/DCseLHr9eKoYT8qsORZK7zVdEGYWHuVtsVrD95wY="], "checkpoint":{"envelope":"rekor.sigstore.dev - 1193050959916656506\n3532873890\nAirPSjrVfD4Np43u5BUpH8irmCFP7voALlh00GvOOek=\n\n— rekor.sigstore.dev wNI9ajBFAiEAzXmjUWULVVaZjfgwukIxuVb4pTataHI5FKbXM/xcj3ECIC39TSyekOQdyyC5/mKqAYMFDBtGO3OEvpnGE+yZv4qt\n"}}, "canonicalizedBody":"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"}], "timestampVerificationData":{"rfc3161Timestamps":[{"signedTimestamp":"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"}]}}, "messageSignature":{"messageDigest":{"algorithm":"SHA2_256", "digest":"Lwzj7ZWW1KEzKzBZ8vB57fAN6xbB5P6D08+RdGlRC4U="}, "signature":"MEUCIGp7nGA4O1Ttyin4NaU6VwZwjPzhlJFXEuuEJp4FJrlnAiEA48KS/o3Hhf+WRUa+bGT3dWLuu9WThbfUT8wcPSYZmxA="}}
|
build/torch-stable-abi29-cu132-x86_64-linux/symbols.json
ADDED
|
@@ -0,0 +1,745 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"schema_version": 1,
|
| 3 |
+
"module": "_flash_attn3_cuda_1ac140b",
|
| 4 |
+
"functions": [
|
| 5 |
+
{
|
| 6 |
+
"name": "add_op_namespace_prefix",
|
| 7 |
+
"qualname": "add_op_namespace_prefix",
|
| 8 |
+
"module": "_flash_attn3_cuda_1ac140b._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_1ac140b.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_1ac140b.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_1ac140b.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_1ac140b.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_1ac140b.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_1ac140b.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-cu134-x86_64-linux/{_flash_attn3_cuda_5fe9dbb.abi3.so → _flash_attn3_cuda_1ac140b.abi3.so}
RENAMED
|
@@ -1,3 +1,3 @@
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
-
oid sha256:
|
| 3 |
size 549439264
|
|
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:b003dbc621087b0634161c477096e0e8cd6201e199078b31591f8a52a5606774
|
| 3 |
size 549439264
|
build/torch-stable-abi29-cu134-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_1ac140b
|
| 3 |
+
ops = torch.ops._flash_attn3_cuda_1ac140b
|
| 4 |
|
| 5 |
def add_op_namespace_prefix(op_name: str):
|
| 6 |
"""
|
| 7 |
Prefix op by namespace.
|
| 8 |
"""
|
| 9 |
+
return f"_flash_attn3_cuda_1ac140b::{op_name}"
|
build/torch-stable-abi29-cu134-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",
|
|
@@ -17,21 +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 |
-
"symbols.json": "
|
| 25 |
}
|
| 26 |
},
|
| 27 |
"provenance": {
|
| 28 |
"kernel-builder": {
|
| 29 |
-
"version": "0.
|
| 30 |
-
"commit": "
|
| 31 |
"dirty": false
|
| 32 |
},
|
| 33 |
"kernel": {
|
| 34 |
-
"commit": "
|
| 35 |
"dirty": false
|
| 36 |
}
|
| 37 |
}
|
|
|
|
| 1 |
{
|
| 2 |
"name": "flash-attn3",
|
| 3 |
+
"id": "_flash_attn3_cuda_1ac140b",
|
| 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_1ac140b.abi3.so": "sAPbxiEIewY0FhxHcJbg6M1iAeGZB4sxWR+KUqVgZ3Q=",
|
| 21 |
+
"_ops.py": "NH6/V6V8u1HinwVJvVwoG30z34Aa9OL/sgYZocQi7a8=",
|
| 22 |
"flash_attn_config.py": "uxo+eyMcDit//8YDaS4Rtk5BTMKzTSK9tOKTjdWVe/w=",
|
| 23 |
"flash_attn_interface.py": "y0bD2JYGMFgX6iDre3zJN2sBPuq/12N6Mo55Z+HQCNQ=",
|
| 24 |
+
"symbols.json": "qzII4ieoZRHwN7dHbNFGa995mlTpWyJs1X5eRdNLU5o="
|
| 25 |
}
|
| 26 |
},
|
| 27 |
"provenance": {
|
| 28 |
"kernel-builder": {
|
| 29 |
+
"version": "0.19.0-dev0",
|
| 30 |
+
"commit": "88e9dd07eb4e639b85799c14b2d8af14a0d9580a",
|
| 31 |
"dirty": false
|
| 32 |
},
|
| 33 |
"kernel": {
|
| 34 |
+
"commit": "1ac140bf36af3c8ff0e73262add55fb8df939e26",
|
| 35 |
"dirty": false
|
| 36 |
}
|
| 37 |
}
|
build/torch-stable-abi29-cu134-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":"3601914180", "logId":{"keyId":"wNI9atQGlz+VWfO6LRygH4QUfY/8W4RFwiT5i5WRgB0="}, "kindVersion":{"kind":"hashedrekord", "version":"0.0.1"}, "integratedTime":"1793392053", "inclusionPromise":{"signedEntryTimestamp":"MEUCIQDBGRp+zjGexppUmbMFrEVSQ3uC+m+GqzDzMURk73aSyQIgJ6yz4JLe0E/2Mglux+5+4jbtllLkYrCJnx/n3KBA0fg="}, "inclusionProof":{"logIndex":"3480009918", "rootHash":"dWatGX974VJRuFN6SpLayXV0G6IvTLfxw4za+pDxQHE=", "treeSize":"3480010141", "hashes":["xBr2NMItoRq1YfI7v9dz08rupqyEOOYz/uKyf7+lSPg=", "UR57Lobs+DCuK2A2zcWRhWlksMBUqICaQ6xf0oDUvQg=", "A5TXo1tsQrAGSKYVoiAt1AhJbwN4F4AO15Yrq1RLiX0=", "b/kvWwPXfoLXirVKpP4D7f5rZal/H0ZRCmjQH1hiLGw=", "vb64WLRMdS3AXtwoWHkZpyIkQOpl9mE3lX82NdRe6iM=", "IIwzO674zmsBLW5yg8u2eFNIZ47kLWvN9asrt2iU0II=", "0PxmCLT+/vohykutKw+D2gBLHu4LOFCtqWrY27x6jdU=", "XltXq9EjBkSA2uluTWlBFSNyvGHBWrktUzYBcgOdGa8=", "0KY3OpaIaOhRKHo75QkiAPzVv6eTwinSRrV4eE000Kw=", "Oyh+2RUa+EstLeFfVhhbFRHFilwlq5s/QfrJMMuvsk4=", "xij9vw5U2ZmpBO0ooE/cTzfl0A8+1ifweHjmiSMLWr8=", "mUtKEp1REU8ipx4txkRb2EFZ4aOx/bgS3JxRL01oG54=", "9t3kgP+u+MFv895vMJiyVlUx0Z9SHSeUhxjsfdwGIos=", "oRoie0t/H+0ZyKRjI6KiDz9stRDDR7EJu7t8qrlVT0Y=", "a8cmhOLBStNzSIVybWzlBd1UxGct2Vy2Eggt5g2B26w=", "39Z7bKKSjPtMpTU8xt8EybgTkCg1qk72XWkQqE5WRSU=", "63Z7Fze4Sm0atCwhCb6EG/N+w2qm+Q+7+dWxf2+uPzA=", "njdy0pCf5uBXEpRPrJGvuVRc6ibqank2dtzoCQ0SR18=", "e3OaWwzL1tEWYXiUzGpH+1v9CiFPwweyfRzIh3Z+uO4=", "b+938HIdfjdD/+LgIs3AVRE3Wc65/1NSSAL4OJ5FUXE=", "kQz/4ND7LdhfAQn+idNsO+ZFC32sfkYRlhjWF5W+sPc=", "zGfeoWKmrZlXMOvyi3R2GDbGpmI1skJW/wLTXCCgF0o=", "1DaDKzrCNNBhmSdY5vFjz+NpAmxmYfYEgBCdPLrMP1A=", "xH/DCseLHr9eKoYT8qsORZK7zVdEGYWHuVtsVrD95wY="], "checkpoint":{"envelope":"rekor.sigstore.dev - 1193050959916656506\n3480010141\ndWatGX974VJRuFN6SpLayXV0G6IvTLfxw4za+pDxQHE=\n\n— rekor.sigstore.dev wNI9ajBFAiAuCu6/AEGAycIStqKp2kS7U3YdR+TpOaXkiaB5X0NAZgIhAIZ8LbNMNNTAxqFGSnnQ+mTQvQtNPMabNovA5CGgzO6Y\n"}}, "canonicalizedBody":"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"}], "timestampVerificationData":{"rfc3161Timestamps":[{"signedTimestamp":"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"}]}}, "messageSignature":{"messageDigest":{"algorithm":"SHA2_256", "digest":"95I2TRNRD+P1lkB8SgwVQqmJ8Rc292MoqdRLftzNjwA="}, "signature":"MEUCIQCnTfMhceXsEsLWcXQZUXZaZSAdt/LMaP6agiPgT4FWmgIgPvCH3TA2JAAMZceLTobDGyEGOA+Wh56oY8kRQym11Jw="}}
|
|
|
|
| 1 |
+
{"mediaType":"application/vnd.dev.sigstore.bundle.v0.3+json", "verificationMaterial":{"certificate":{"rawBytes":"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"}, "tlogEntries":[{"logIndex":"3654778576", "logId":{"keyId":"wNI9atQGlz+VWfO6LRygH4QUfY/8W4RFwiT5i5WRgB0="}, "kindVersion":{"kind":"hashedrekord", "version":"0.0.1"}, "integratedTime":"1793669016", "inclusionPromise":{"signedEntryTimestamp":"MEYCIQDD/73stcw0aoxSSZsegkyMdWhF9SF4BhaAkmGsmwcUjAIhAOvJnNWk8XfbA6eMirpZsUK4iDMIq1WoUEEw9Usc4qVo"}, "inclusionProof":{"logIndex":"3532874314", "rootHash":"8dhU5ELb4ogHTLBfISzu7ZWDT0tAnm+YGUeDw4rifOk=", "treeSize":"3532874350", "hashes":["vHH6vf6xEBrXu0EQZXVwVoHYyTJZyH5JFc9Me5G5m/w=", "HqdG1y82GMWaxiiQRJwsdcdtcYHJCyfEDBRv0p4dBZU=", "V0FIEGzFUjmJtk9wXLrDvmSBYVe+zCbANZr6H9G8kbU=", "4aNY9XeWK98YiSah+2Cag399z6PpZ8KmgpXBXklkito=", "77tmCFPP+fTLNWi8HLbza7b1p+r2uimVplJoeHysNTQ=", "75ZkKLhW0FG+Vk1euu0r/s+gb+DNvHvScD4HXxeoEiU=", "qQflVDJl/2OBu711UyV8ht9T0Vqbuj3cpWGyJkHQb1k=", "E0EEpZv4h1OdmyLYUDloW4BGHXc2Z/VRPhtdPDGllpg=", "VlrHGUxNk5PYgNCjDyZFwri2UY985FoN1fBJ2z6y/Aw=", "jXRnHZhdIEDN/dOSOGnahiPHiCL8svXXZkNlpklSSDE=", "JgdHnyIJAAm3PA7mCUWsdLQ/ZPPYPC6Oh1Gw79WHtpg=", "MlqLoiFm8Oo5bbmftEQbgcSKXfa+i5FCiungGzOC1uo=", "QYcxedNInCRxBCNHkv8RT2T790mUNX70dZsQ78gLIYQ=", "ogjRZ88pnEW/ANOWeLXfhYL/XdgeGJy0zJTgCayJlGE=", "PA9MhoiExYPp/vXh8tVCBrATkoY31KStgze+rovDTkA=", "SlixnS34BhAO+yRI2FaGPSOh9V6JSVNPfLl1bHxrcQY=", "1DaDKzrCNNBhmSdY5vFjz+NpAmxmYfYEgBCdPLrMP1A=", "xH/DCseLHr9eKoYT8qsORZK7zVdEGYWHuVtsVrD95wY="], "checkpoint":{"envelope":"rekor.sigstore.dev - 1193050959916656506\n3532874350\n8dhU5ELb4ogHTLBfISzu7ZWDT0tAnm+YGUeDw4rifOk=\n\n— rekor.sigstore.dev wNI9ajBGAiEA0hwLgdzAzppyk4PUdj+CSeRo/OdQvbVo9BZXpEFsfRwCIQCGWDO+oCPLaVRhexa+X5DnWlBJO96Qin/u44AMOGSDrg==\n"}}, "canonicalizedBody":"eyJhcGlWZXJzaW9uIjoiMC4wLjEiLCJraW5kIjoiaGFzaGVkcmVrb3JkIiwic3BlYyI6eyJkYXRhIjp7Imhhc2giOnsiYWxnb3JpdGhtIjoic2hhMjU2IiwidmFsdWUiOiI0N2FiMjFiZjI1ZTJmM2NhNjUyY2Q3YjQ4ZGFkN2YwOGVhMTI2ZDkzZGEzYWZiYjdhMWJlYjc0M2YzYWVkMjhkIn19LCJzaWduYXR1cmUiOnsiY29udGVudCI6Ik1FVUNJUUN3S0pKR21DeXFQWHNxVWZac0ZOQmYyNzI1ZnhySFNQWlJZZzcxNWdGdnlRSWdjTUQyL3IwSnJKSVB1akF5Z2dYWkliSU9MQjZWdlVYdEY1SktGZGZRMDU0PSIsInB1YmxpY0tleSI6eyJjb250ZW50IjoiTFMwdExTMUNSVWRKVGlCRFJWSlVTVVpKUTBGVVJTMHRMUzB0Q2sxSlNVaE1SRU5EUW5KUFowRjNTVUpCWjBsVlkwUnFSbFZxUzBGaE1tc3dkbGcwVEhGTVRsaEhMek0yZG5GQmQwTm5XVWxMYjFwSmVtb3dSVUYzVFhjS1RucEZWazFDVFVkQk1WVkZRMmhOVFdNeWJHNWpNMUoyWTIxVmRWcEhWakpOVWpSM1NFRlpSRlpSVVVSRmVGWjZZVmRrZW1SSE9YbGFVekZ3WW01U2JBcGpiVEZzV2tkc2FHUkhWWGRJYUdOT1RXcFplRTFVUVhwTlJFVjVUWHBOTWxkb1kwNU5hbGw0VFZSQmVrMUVSWHBOZWsweVYycEJRVTFHYTNkRmQxbElDa3R2V2tsNmFqQkRRVkZaU1V0dldrbDZhakJFUVZGalJGRm5RVVZvWTI5aWVHSnZUSEp5YUZobE5XZHlabGxyYjBGUlRHczFUR3BzVjJaS2JHNU9aM2tLU1ZoQ1ZVcG5VVTByTUd3eWRXVmxVbVZUZEROV2JGTnVNMlZWUlcxclZqZDJkMGh3TjFaa05YRnNPRFJXU0dwUlpIRlBRMEprU1hkbloxaFBUVUUwUndwQk1WVmtSSGRGUWk5M1VVVkJkMGxJWjBSQlZFSm5UbFpJVTFWRlJFUkJTMEpuWjNKQ1owVkdRbEZqUkVGNlFXWkNaMDVXU0ZOTlJVZEVRVmRuUWxSbUNqQXJibEJXYVZGU2JIWnRiekpQYTI5V1lVeEhUR2hvYTFCNlFuSkNaMDVXU0ZKRlFrRm1PRVZaVkVKbWFHd3hiMlJJVW5kamVtOTJUREprY0dSSGFERUtXV2sxYW1JeU1IWmhTRlp1V2pKc2RWb3lXbWhaTWxWMllUSldlV0p0Vm5OamVURnFZakl4ZEdSWE5YQmtTR3QyVEcxa2NHUkhhREZaYVRrellqTktjZ3BhYlhoMlpETk5kbGx1Vm5CaVIxRjFaVmRHZEdKRlFubGFWMXA2VERKb2JGbFhVbnBNTWpGb1lWYzBkMDlSV1V0TGQxbENRa0ZIUkhaNlFVSkJVVkZ5Q21GSVVqQmpTRTAyVEhrNU1HSXlkR3hpYVRWb1dUTlNjR0l5TlhwTWJXUndaRWRvTVZsdVZucGFXRXBxWWpJMU1GcFhOVEJNYlU1MllsUkJaa0puYjNJS1FtZEZSVUZaVHk5TlFVVkRRa0pHTTJJelNuSmFiWGgyWkRFNWEyRllUbmRaV0ZKcVlVUkJNa0puYjNKQ1owVkZRVmxQTDAxQlJVUkNRMmQ0V1ZkTmVBcE9SRUpwV21wTk1sbFhXWHBaZW1odFdtcENiRTU2VFhsT2FrcG9Xa2RSTVU1WFdtbFBSMUp0VDFSTk5WcFVTVEpOUWsxSFEybHpSMEZSVVVKbk56aDNDa0ZSVVVWQ1ZVb3hZVmQ0YTAxRGMwZERhWE5IUVZGUlFtYzNPSGRCVVZWRlNGZG9NVm95WkhCaWJXUnRXVmRPYkV3eWRHeGpiVFZzWWtoTmRGa3lPWFFLWWxoV2RXRllValZOUWpCSFEybHpSMEZSVVVKbk56aDNRVkZaUlVRelNteGFiazEyWVVkV2FGcElUWFppVjBad1ltcEJOMEpuYjNKQ1owVkZRVmxQTHdwTlFVVkpRa013VFVzeWFEQmtTRUo2VDJrNGRtUkhPWEphVnpSMVdWZE9NR0ZYT1hWamVUVnVZVmhTYjJSWFNqRmpNbFo1V1RJNWRXUkhWblZrUXpWcUNtSXlNSGRpVVZsTFMzZFpRa0pCUjBSMmVrRkNRMUZTWmtSR01XOWtTRkozWTNwdmRrd3laSEJrUjJneFdXazFhbUl5TUhaaFNGWnVXakpzZFZveVdtZ0tXVEpWZG1FeVZubGliVlp6WTNreGFtSXlNWFJrVnpWd1pFaHJka3h0WkhCa1IyZ3hXV2s1TTJJelNuSmFiWGgyWkROTmRsbHVWbkJpUjFGMVpWZEdkQXBpUlVKNVdsZGFla3d5YUd4WlYxSjZUREl4YUdGWE5IZFBRVmxMUzNkWlFrSkJSMFIyZWtGQ1EyZFJjVVJEWjNoWlYwMTRUa1JDYVZwcVRUSlpWMWw2Q2xsNmFHMWFha0pzVG5wTmVVNXFTbWhhUjFFeFRsZGFhVTlIVW0xUFZFMDFXbFJKTWsxQ2MwZERhWE5IUVZGUlFtYzNPSGRCVVhORlJGRjNUR015Vm5NS1dta3hiMkl6VGpCYVYxRjNVVUZaUzB0M1dVSkNRVWRFZG5wQlFrUkJVWGxFUkVKdlpFaFNkMk42YjNaTU1tUndaRWRvTVZscE5XcGlNakIyWVVoV2JncGFNbXgxV2pKYWFGa3lWWFpoTWxaNVltMVdjMk41TVdwaU1qRjBaRmMxY0dSSWEzZFBRVmxMUzNkWlFrSkJSMFIyZWtGQ1JGRlJjVVJEWjNoWlYwMTRDazVFUW1sYWFrMHlXVmRaZWxsNmFHMWFha0pzVG5wTmVVNXFTbWhhUjFFeFRsZGFhVTlIVW0xUFZFMDFXbFJKTWsxQ09FZERhWE5IUVZGUlFtYzNPSGNLUVZFMFJVVlJkMUJqYlZadFkzazViMXBYUm10amVUbDBXVmRzZFUxQ2IwZERhWE5IUVZGUlFtYzNPSGRCVVRoRlJFRjNTMDFVUVROTlZGRXpUbFJWZVFwUFZFRjFRbWR2Y2tKblJVVkJXVTh2VFVGRlVVSkRRVTFJYldnd1pFaENlazlwT0haYU1td3dZVWhXYVV4dFRuWmlVemx2WkZka2JtRlhOVzVhYlVacUNscFVRVmxDWjI5eVFtZEZSVUZaVHk5TlFVVlNRa0Z2VFVORVNURk9la2wzVG5wUmVrMUhNRWREYVhOSFFWRlJRbWMzT0hkQlVrbEZXSGQ0WkdGSVVqQUtZMGhOTmt4NU9XNWhXRkp2WkZkSmRWa3lPWFJNTW1neFdqSmtjR0p0WkcxWlYwNXNUREowYkdOdE5XeGlTRTEwV1RJNWRHSllWblZoV0ZJMVRIazFiZ3BoV0ZKdlpGZEpkbVF5T1hsaE1scHpZak5rZWt3eVNqRmhWM2hyVEc1c2FHSlhlRUZqYlZadFkzazViMXBYUm10amVUbDBXVmRzZFUxRVowZERhWE5IQ2tGUlVVSm5OemgzUVZKTlJVdG5kMjlOVjBacVRWUlJkMWx0V1hwT2JVWnRUVEpOTkZwdFdYZGFWR042VFdwWmVWbFhVbXRPVkZadFdXcG9hMXBxYTNvS1QxZFZlVTVxUVdoQ1oyOXlRbWRGUlVGWlR5OU5RVVZWUWtKTlRVVllaSFpqYlhSdFlrYzVNMWd5VW5Cak0wSm9aRWRPYjAxSFVVZERhWE5IUVZGUlFncG5OemgzUVZKVlJWWm5lRlZoU0ZJd1kwaE5Oa3g1T1c1aFdGSnZaRmRKZFZreU9YUk1NbWd4V2pKa2NHSnRaRzFaVjA1c1RESjBiR050Tld4aVNFMTBDbGt5T1hSaVdGWjFZVmhTTlV3eVJtcGtSMngyWW01TmRtTnVWblZqZVRnd1RWUlZkMDlVWnpOT2VsRXhUV2s1YUdSSVVteGlXRUl3WTNrNGVFMUNXVWNLUTJselIwRlJVVUpuTnpoM1FWSlpSVU5CZDBkalNGWnBZa2RzYWsxRldVZERhWE5IUVZGUlFtYzNPSGRCVW1kRlQwRjNNbU50Vm5kaWVuQnZaRmRrYmdwaFZ6VnVXbTFHYWxwVE9YSmFXRXAxV2xkNGVreFhUblppVnpFeFltMXNNR1ZVY0hsYVYxazJZMjFXYldONU9XOWFWMFpyWTNrNWRGbFhiSFZOU1VkTUNrSm5iM0pDWjBWRlFXUmFOVUZuVVVOQ1NEQkZaWGRDTlVGSVkwRXpWREIzWVhOaVNFVlVTbXBIVWpSamJWZGpNMEZ4U2t0WWNtcGxVRXN6TDJnMGNIa0taME00Y0Rkdk5FRkJRVWRvYm5kRmNuTlJRVUZDUVUxQlUwUkNSMEZwUlVGcGJIQjBRMmh0WkhoVk9WRm9lREUxT1hSM1F6ZDFibkZQTkRnemRFUjRPQXBEVUVoME16bDFiSGRTWjBOSlVVTm9VMkpVVlcwMlJXcEdMMGxsVWprMmFtZDZhVE5YWlVkblNqWnZaemxZTW5Sbk9GSk5iekp3Vm1oRVFVdENaMmR4Q21ocmFrOVFVVkZFUVhkT2JrRkVRbXRCYWtGVWJUSndUVmQyYXpOWWIzTjJNVWgxT0VsdVlURjJaVkZyYlZoS1dYaDFkMjFIUW14U1dTdENjRm95TlRBS2NVNTVLMmxhY21GMVRVMXBaM1psVFRKYWQwTk5RamN3UlVSQllVdHNTMkpVZEVaVlJ5OUlhMGd3TWsxb2NEVjVUbHBUUWxkTWRFbE9TVXhhZGtWelR3cHhXWEJIV0hab1YzcDZhMGN2YUdoTUszRTNVM1JCUFQwS0xTMHRMUzFGVGtRZ1EwVlNWRWxHU1VOQlZFVXRMUzB0TFFvPSJ9fX19"}], "timestampVerificationData":{"rfc3161Timestamps":[{"signedTimestamp":"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"}]}}, "messageSignature":{"messageDigest":{"algorithm":"SHA2_256", "digest":"R6shvyXi88plLNe0ja1/COoSbZPaOvu3ob63Q/Ou0o0="}, "signature":"MEUCIQCwKJJGmCyqPXsqUfZsFNBf2725fxrHSPZRYg715gFvyQIgcMD2/r0JrJIPujAyggXZIbIOLB6VvUXtF5JKFdfQ054="}}
|
build/torch-stable-abi29-cu134-x86_64-linux/symbols.json
CHANGED
|
@@ -1,11 +1,11 @@
|
|
| 1 |
{
|
| 2 |
"schema_version": 1,
|
| 3 |
-
"module": "
|
| 4 |
"functions": [
|
| 5 |
{
|
| 6 |
"name": "add_op_namespace_prefix",
|
| 7 |
"qualname": "add_op_namespace_prefix",
|
| 8 |
-
"module": "
|
| 9 |
"kind": "function",
|
| 10 |
"doc": "Prefix op by namespace.",
|
| 11 |
"signature": {
|
|
@@ -23,7 +23,7 @@
|
|
| 23 |
{
|
| 24 |
"name": "flash_attn_combine",
|
| 25 |
"qualname": "flash_attn_combine",
|
| 26 |
-
"module": "
|
| 27 |
"kind": "function",
|
| 28 |
"doc": null,
|
| 29 |
"signature": {
|
|
@@ -59,7 +59,7 @@
|
|
| 59 |
{
|
| 60 |
"name": "flash_attn_func",
|
| 61 |
"qualname": "flash_attn_func",
|
| 62 |
-
"module": "
|
| 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": {
|
|
@@ -173,7 +173,7 @@
|
|
| 173 |
{
|
| 174 |
"name": "flash_attn_qkvpacked_func",
|
| 175 |
"qualname": "flash_attn_qkvpacked_func",
|
| 176 |
-
"module": "
|
| 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": {
|
|
@@ -263,7 +263,7 @@
|
|
| 263 |
{
|
| 264 |
"name": "flash_attn_varlen_func",
|
| 265 |
"qualname": "flash_attn_varlen_func",
|
| 266 |
-
"module": "
|
| 267 |
"kind": "function",
|
| 268 |
"doc": null,
|
| 269 |
"signature": {
|
|
@@ -413,7 +413,7 @@
|
|
| 413 |
{
|
| 414 |
"name": "flash_attn_with_kvcache",
|
| 415 |
"qualname": "flash_attn_with_kvcache",
|
| 416 |
-
"module": "
|
| 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": {
|
|
@@ -605,7 +605,7 @@
|
|
| 605 |
{
|
| 606 |
"name": "get_scheduler_metadata",
|
| 607 |
"qualname": "get_scheduler_metadata",
|
| 608 |
-
"module": "
|
| 609 |
"kind": "function",
|
| 610 |
"doc": null,
|
| 611 |
"signature": {
|
|
|
|
| 1 |
{
|
| 2 |
"schema_version": 1,
|
| 3 |
+
"module": "_flash_attn3_cuda_1ac140b",
|
| 4 |
"functions": [
|
| 5 |
{
|
| 6 |
"name": "add_op_namespace_prefix",
|
| 7 |
"qualname": "add_op_namespace_prefix",
|
| 8 |
+
"module": "_flash_attn3_cuda_1ac140b._ops",
|
| 9 |
"kind": "function",
|
| 10 |
"doc": "Prefix op by namespace.",
|
| 11 |
"signature": {
|
|
|
|
| 23 |
{
|
| 24 |
"name": "flash_attn_combine",
|
| 25 |
"qualname": "flash_attn_combine",
|
| 26 |
+
"module": "_flash_attn3_cuda_1ac140b.flash_attn_interface",
|
| 27 |
"kind": "function",
|
| 28 |
"doc": null,
|
| 29 |
"signature": {
|
|
|
|
| 59 |
{
|
| 60 |
"name": "flash_attn_func",
|
| 61 |
"qualname": "flash_attn_func",
|
| 62 |
+
"module": "_flash_attn3_cuda_1ac140b.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": {
|
|
|
|
| 173 |
{
|
| 174 |
"name": "flash_attn_qkvpacked_func",
|
| 175 |
"qualname": "flash_attn_qkvpacked_func",
|
| 176 |
+
"module": "_flash_attn3_cuda_1ac140b.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": {
|
|
|
|
| 263 |
{
|
| 264 |
"name": "flash_attn_varlen_func",
|
| 265 |
"qualname": "flash_attn_varlen_func",
|
| 266 |
+
"module": "_flash_attn3_cuda_1ac140b.flash_attn_interface",
|
| 267 |
"kind": "function",
|
| 268 |
"doc": null,
|
| 269 |
"signature": {
|
|
|
|
| 413 |
{
|
| 414 |
"name": "flash_attn_with_kvcache",
|
| 415 |
"qualname": "flash_attn_with_kvcache",
|
| 416 |
+
"module": "_flash_attn3_cuda_1ac140b.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": {
|
|
|
|
| 605 |
{
|
| 606 |
"name": "get_scheduler_metadata",
|
| 607 |
"qualname": "get_scheduler_metadata",
|
| 608 |
+
"module": "_flash_attn3_cuda_1ac140b.flash_attn_interface",
|
| 609 |
"kind": "function",
|
| 610 |
"doc": null,
|
| 611 |
"signature": {
|