SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2407.08608 · 2024

FlashAttention-3: Fast and Accurate Attention with Asynchrony and Low-precision

arXiv · PDF · Open in the Atlas

Code that ran

16 of 18 samples (functions and classes) lifted from this paper's linked repositories executed on a synthesized input in Syntology's jail (0 of them constructed an object rather than computing a result; 5 honoured the contract we drafted, 0 violated it and 0 have no contract verdict; 11 failures of our instrument; 2 no recorded run), measured 2026-11-22; repository links from the Papers with Code archive attributed under CC BY-SA, no affiliation; bodies inline only under a permissive licence as GitHub detects it. Not a reproduction of the paper's results.

"Ran" means the function executed on a synthesized input and returned a value. It is not a reproduction of the paper's results. This record, dated, as JSON: /api/ran/2407.08608.json · badge: /api/ran/2407.08608.svg

Each repository's role is a label: a classifier's output recorded on the link, or, where no role is recorded, what the Papers with Code archive marked. The line under it names its basis. Neither says who wrote the code.

RepositoryRoleRan
dao-ailab/flash-attention canonical (classifier label)
classifier basis: census
16 ran · 2 no recorded run
FunctionStatusWhere it lives
apply_rotary_emb_torch Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
dao-ailab/flash-attention/flash_attn/layers/rotary.py
pointer only (licence: BSD-3-Clause); the stored copy get_code serves, from amazon-science/dualkv-flash-attn-for-rl, is pointer only (licence: NOASSERTION) · get_code("5ec949977c334f87")
How it ran

The body of this sample is withheld: this paper's link allows the body inline (BSD-3-Clause), but get_code serves the stored copy, first harvested from amazon-science/dualkv-flash-attn-for-rl, and GitHub could not classify the licence of that repository (NOASSERTION). The setup block that called it is withheld with it, because it is written against that body. source_path and origin say where the code lives, to read under the repository's own terms. A run is recorded.

attention_pytorch Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
dao-ailab/flash-attention/flash_attn/cute/benchmark_flash_attention_fp8.py
code served (permissive licence) · get_code("4afe00a5853d633f")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1) and returned Tensor, ndim 4, dtype family float, shape (2, 4, 3, 4) as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import math

torch.manual_seed(0)

# Create a small qkv tensor
# Shape: (batch, seqlen, 3, nheads, headdim)
batch_size = 2
seqlen = 4
nheads = 3
headdim = 4

qkv = torch.randn(batch_size, seqlen, 3, nheads, headdim, dtype=torch.float32)

# causal is a boolean flag
causal = True

ARGS = {
    "qkv": qkv,
    "causal": causal,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1
Imports
einops, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Tensor, ndim 4, dtype family float, shape (2, 4, 3, 4) — shape (2, 4, 3, 4) as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
attention_pytorch Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
dao-ailab/flash-attention/benchmarks/benchmark_flash_attention.py
code served (permissive licence) · get_code("942ce4d7b0714c0a")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned Tensor, ndim 4, dtype family float, shape (2, 4, 3, 8) as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import math
import torch
import torch.nn.functional as F
from einops import rearrange, repeat

torch.manual_seed(0)

# Construct a minimal qkv tensor
# Shape: (batch_size, seqlen, 3, nheads, head_dim)
batch_size = 2
seqlen = 4
nheads = 3
head_dim = 8

qkv = torch.randn(batch_size, seqlen, 3, nheads, head_dim, dtype=torch.float32)

ARGS = {
    "qkv": qkv,
    "dropout_p": 0.0,
    "causal": True,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
einops, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Tensor, ndim 4, dtype family float, shape (2, 4, 3, 8) — shape (2, 4, 3, 8) as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
convert_blockmask Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
dao-ailab/flash-attention/flash_attn/flash_blocksparse_attn_interface.py
pointer only (licence: BSD-3-Clause); the stored copy get_code serves, from amazon-science/dualkv-flash-attn-for-rl, is pointer only (licence: NOASSERTION) · get_code("cc628b160e090ba1")
How it ran

The body of this sample is withheld: this paper's link allows the body inline (BSD-3-Clause), but get_code serves the stored copy, first harvested from amazon-science/dualkv-flash-attn-for-rl, and GitHub could not classify the licence of that repository (NOASSERTION). The setup block that called it is withheld with it, because it is written against that body. source_path and origin say where the code lives, to read under the repository's own terms. A run is recorded.

efficiency Ran
ran, honoured the contract we drafted
dao-ailab/flash-attention/flash_attn/cute/benchmark_flash_attention_fp8.py
code served (permissive licence) · get_code("f29cb637db3ec374")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1) and returned float, ndim 0, dtype family float, shape () as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import math

def efficiency(flop: int, seconds: float) -> float:
    return flop / seconds / 1000000000000.0 if not math.isnan(seconds) else 0.0

ARGS = {
    "flop": 1000000000000,
    "seconds": 2.5
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type float, ndim 0, dtype family float, shape () — shape () as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
efficiency Ran
ran, honoured the contract we drafted
dao-ailab/flash-attention/benchmarks/benchmark_flash_attention.py
code served (permissive licence) · get_code("b5de75d142329f03")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned float, ndim 0, dtype family float, shape () as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import math
import torch

torch.manual_seed(0)

# The function efficiency(flop, time) computes flop / time / 10**12
# if time is not NaN, else returns 0.0
# Both parameters are numeric values

flop_value = 1e12  # 1 trillion FLOPs
time_value = 2.5   # 2.5 seconds

ARGS = {
    "flop": flop_value,
    "time": time_value,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type float, ndim 0, dtype family float, shape () — shape () as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
flash_attn_qkvpacked_func Ran
ran; our contract draft was wrong, not the code
dao-ailab/flash-attention/flash_attn/flash_attn_interface.py
pointer only (licence: BSD-3-Clause); the stored copy get_code serves, from amazon-science/dualkv-flash-attn-for-rl, is pointer only (licence: NOASSERTION) · get_code("7541ce9614862680")
How it ran

The body of this sample is withheld: this paper's link allows the body inline (BSD-3-Clause), but get_code serves the stored copy, first harvested from amazon-science/dualkv-flash-attn-for-rl, and GitHub could not classify the licence of that repository (NOASSERTION). The setup block that called it is withheld with it, because it is written against that body. source_path and origin say where the code lives, to read under the repository's own terms. A run is recorded.

flops Ran
ran, honoured the contract we drafted
dao-ailab/flash-attention/flash_attn/cute/benchmark_flash_attention_fp8.py
code served (permissive licence) · get_code("e73bdfbda28da341")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1) and returned int, ndim 0, dtype family int, shape () as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import torch.nn as nn

torch.manual_seed(0)

ARGS = {
    "batch": 2,
    "seqlen": 4,
    "headdim": 3,
    "nheads": 5,
    "causal": True,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type int, ndim 0, dtype family int, shape () — shape () as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
flops Ran
ran, honoured the contract we drafted
dao-ailab/flash-attention/benchmarks/benchmark_flash_attention.py
code served (permissive licence) · get_code("8d5fb28ca2f46eb8")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned int, ndim 0, dtype family int, shape () as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch

torch.manual_seed(0)

ARGS = {
    'batch': 2,
    'seqlen': 4,
    'headdim': 8,
    'nheads': 3,
    'causal': True,
    'mode': 'fwd'
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type int, ndim 0, dtype family int, shape () — shape () as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
init_to_zero Ran
ran; our contract draft was wrong, not the code
dao-ailab/flash-attention/flash_attn/flash_attn_triton.py
pointer only (licence: BSD-3-Clause); the stored copy get_code serves, from amazon-science/dualkv-flash-attn-for-rl, is pointer only (licence: NOASSERTION) · get_code("1586b639a5abd0b0")
How it ran

The body of this sample is withheld: this paper's link allows the body inline (BSD-3-Clause), but get_code serves the stored copy, first harvested from amazon-science/dualkv-flash-attn-for-rl, and GitHub could not classify the licence of that repository (NOASSERTION). The setup block that called it is withheld with it, because it is written against that body. source_path and origin say where the code lives, to read under the repository's own terms. A run is recorded.

maybe_contiguous Ran
ran; our contract draft was wrong, not the code
dao-ailab/flash-attention/flash_attn/flash_attn_interface.py
pointer only (licence: BSD-3-Clause); the stored copy get_code serves, from amazon-science/dualkv-flash-attn-for-rl, is pointer only (licence: NOASSERTION) · get_code("19ad2ccd82e378c1")
How it ran

The body of this sample is withheld: this paper's link allows the body inline (BSD-3-Clause), but get_code serves the stored copy, first harvested from amazon-science/dualkv-flash-attn-for-rl, and GitHub could not classify the licence of that repository (NOASSERTION). The setup block that called it is withheld with it, because it is written against that body. source_path and origin say where the code lives, to read under the repository's own terms. A run is recorded.

pad_input Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
dao-ailab/flash-attention/flash_attn/bert_padding.py
pointer only (licence: BSD-3-Clause); the stored copy get_code serves, from amazon-science/dualkv-flash-attn-for-rl, is pointer only (licence: NOASSERTION) · get_code("a8cf6124094beaab")
How it ran

The body of this sample is withheld: this paper's link allows the body inline (BSD-3-Clause), but get_code serves the stored copy, first harvested from amazon-science/dualkv-flash-attn-for-rl, and GitHub could not classify the licence of that repository (NOASSERTION). The setup block that called it is withheld with it, because it is written against that body. source_path and origin say where the code lives, to read under the repository's own terms. A run is recorded.

rotate_half Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
dao-ailab/flash-attention/flash_attn/layers/rotary.py
pointer only (licence: BSD-3-Clause); the stored copy get_code serves, from amazon-science/dualkv-flash-attn-for-rl, is pointer only (licence: NOASSERTION) · get_code("163292f43a3c861b")
How it ran

The body of this sample is withheld: this paper's link allows the body inline (BSD-3-Clause), but get_code serves the stored copy, first harvested from amazon-science/dualkv-flash-attn-for-rl, and GitHub could not classify the licence of that repository (NOASSERTION). The setup block that called it is withheld with it, because it is written against that body. source_path and origin say where the code lives, to read under the repository's own terms. A run is recorded.

round_multiple Ran
ran, honoured the contract we drafted
dao-ailab/flash-attention/flash_attn/flash_attn_interface.py
pointer only (licence: BSD-3-Clause); the stored copy get_code serves, from amazon-science/dualkv-flash-attn-for-rl, is pointer only (licence: NOASSERTION) · get_code("a0595f1c95cacf06")
How it ran

The body of this sample is withheld: this paper's link allows the body inline (BSD-3-Clause), but get_code serves the stored copy, first harvested from amazon-science/dualkv-flash-attn-for-rl, and GitHub could not classify the licence of that repository (NOASSERTION). The setup block that called it is withheld with it, because it is written against that body. source_path and origin say where the code lives, to read under the repository's own terms. A run is recorded.

unpad_input Ran
ran; our contract draft was wrong, not the code
dao-ailab/flash-attention/flash_attn/bert_padding.py
pointer only (licence: BSD-3-Clause); the stored copy get_code serves, from amazon-science/dualkv-flash-attn-for-rl, is pointer only (licence: NOASSERTION) · get_code("d84093c20a0ed8fd")
How it ran

The body of this sample is withheld: this paper's link allows the body inline (BSD-3-Clause), but get_code serves the stored copy, first harvested from amazon-science/dualkv-flash-attn-for-rl, and GitHub could not classify the licence of that repository (NOASSERTION). The setup block that called it is withheld with it, because it is written against that body. source_path and origin say where the code lives, to read under the repository's own terms. A run is recorded.

unpad_input_for_concatenated_sequences Ran
ran; our contract draft was wrong, not the code
dao-ailab/flash-attention/flash_attn/bert_padding.py
pointer only (licence: BSD-3-Clause); the stored copy get_code serves, from amazon-science/dualkv-flash-attn-for-rl, is pointer only (licence: NOASSERTION) · get_code("62394c842aafb022")
How it ran

The body of this sample is withheld: this paper's link allows the body inline (BSD-3-Clause), but get_code serves the stored copy, first harvested from amazon-science/dualkv-flash-attn-for-rl, and GitHub could not classify the licence of that repository (NOASSERTION). The setup block that called it is withheld with it, because it is written against that body. source_path and origin say where the code lives, to read under the repository's own terms. A run is recorded.

convert_to_cudnn_type Not yet run
unverified (harvested, no recorded run)
dao-ailab/flash-attention/hopper/benchmark_flash_attention_fp8.py
code served (permissive licence) · get_code("a9519adef33152bb")
How it ran

no invocation recorded.

cudnn_spda_setup Not yet run
unverified (harvested, no recorded run)
dao-ailab/flash-attention/hopper/benchmark_flash_attention_fp8.py
code served (permissive licence) · get_code("58ebcd8dbefa010d")
How it ran

no invocation recorded.

Also present in other papers' code

Of 18 samples lifted from this paper's linked repositories, 18 carry a function identity and 0 do not. Of the 18 distinct functions with an identity, 18 are also present in other papers whose code we hold (18 with at least one paper dated earlier, 1 the same month, 18 later) and 0 are only here among the 52,977 papers whose code Syntology holds (as of 2026-09-26).

Only here means: among the 52,977 papers whose code Syntology holds (51,887 with lifted samples, 1,090 through their assemblies only), as of 2026-09-26, no other paper's lifted samples or current assembly bricks contain an identical function body after normalisation.

FunctionAlso present inEarliest-datedOthers named
init_to_zero 12 other papers
7 dated earlier, 2 the same month, 3 later
2304.11277 (2023-04)
PyTorch FSDP: Experiences on Scaling Fully Sharded Data Parallel
2309.06180 (2023-09)
Efficient Memory Management for Large Language Model Serving with PagedAttention
2310.05869 (2023-10)
HyperAttention: Long-context Attention in Near-Linear Time
and 9 more not named here
apply_rotary_emb_torch 6 other papers
1 dated earlier, 5 later
2309.06180 (2023-09)
Efficient Memory Management for Large Language Model Serving with PagedAttention
2412.06464 (2024-12)
Gated Delta Networks: Improving Mamba2 with Delta Rule
2601.17367 (2026-01)
Elastic Attention: Test-time Adaptive Sparsity Ratios for Efficient Transformers
and 3 more not named here
rotate_half 6 other papers
1 dated earlier, 5 later
2309.06180 (2023-09)
Efficient Memory Management for Large Language Model Serving with PagedAttention
2412.06464 (2024-12)
Gated Delta Networks: Improving Mamba2 with Delta Rule
2511.10056 (2025-11)
From Static Structures to Ensembles: Studying and Harnessing Protein Structure Tokenization
and 3 more not named here
pad_input 4 other papers
1 dated earlier, 3 later
2309.06180 (2023-09)
Efficient Memory Management for Large Language Model Serving with PagedAttention
2502.19613 (2025-02)
Self-rewarding correction for mathematical reasoning
2604.23798 (2026-04)
ELSA: Exact Linear-Scan Attention for Fast and Memory-Light Vision Transformers
and 1 more not named here
unpad_input 4 other papers
1 dated earlier, 3 later
2309.06180 (2023-09)
Efficient Memory Management for Large Language Model Serving with PagedAttention
2502.19613 (2025-02)
Self-rewarding correction for mathematical reasoning
2604.23798 (2026-04)
ELSA: Exact Linear-Scan Attention for Fast and Memory-Light Vision Transformers
and 1 more not named here
unpad_input_for_concatenated_sequences 4 other papers
1 dated earlier, 3 later
2309.06180 (2023-09)
Efficient Memory Management for Large Language Model Serving with PagedAttention
2502.19613 (2025-02)
Self-rewarding correction for mathematical reasoning
2604.23798 (2026-04)
ELSA: Exact Linear-Scan Attention for Fast and Memory-Light Vision Transformers
and 1 more not named here
attention_pytorch 3 other papers
2 dated earlier, 1 later
2309.06180 (2023-09)
Efficient Memory Management for Large Language Model Serving with PagedAttention
2312.07104 (2023-12)
SGLang: Efficient Execution of Structured Language Model Programs
2604.23798 (2026-04)
ELSA: Exact Linear-Scan Attention for Fast and Memory-Light Vision Transformers
convert_blockmask 3 other papers
1 dated earlier, 2 later
2309.06180 (2023-09)
Efficient Memory Management for Large Language Model Serving with PagedAttention
2604.23798 (2026-04)
ELSA: Exact Linear-Scan Attention for Fast and Memory-Light Vision Transformers
2605.15422 (2026-05)
DualKV: Shared-Prompt Flash Attention for Efficient RL Training with Large Rollouts and Long Contexts
efficiency 3 other papers
2 dated earlier, 1 later
2309.06180 (2023-09)
Efficient Memory Management for Large Language Model Serving with PagedAttention
2312.07104 (2023-12)
SGLang: Efficient Execution of Structured Language Model Programs
2604.23798 (2026-04)
ELSA: Exact Linear-Scan Attention for Fast and Memory-Light Vision Transformers
flash_attn_qkvpacked_func 3 other papers
1 dated earlier, 2 later
2309.06180 (2023-09)
Efficient Memory Management for Large Language Model Serving with PagedAttention
2604.23798 (2026-04)
ELSA: Exact Linear-Scan Attention for Fast and Memory-Light Vision Transformers
2605.15422 (2026-05)
DualKV: Shared-Prompt Flash Attention for Efficient RL Training with Large Rollouts and Long Contexts
flops 3 other papers
2 dated earlier, 1 later
2309.06180 (2023-09)
Efficient Memory Management for Large Language Model Serving with PagedAttention
2312.07104 (2023-12)
SGLang: Efficient Execution of Structured Language Model Programs
2604.23798 (2026-04)
ELSA: Exact Linear-Scan Attention for Fast and Memory-Light Vision Transformers
maybe_contiguous 3 other papers
1 dated earlier, 2 later
2309.06180 (2023-09)
Efficient Memory Management for Large Language Model Serving with PagedAttention
2604.23798 (2026-04)
ELSA: Exact Linear-Scan Attention for Fast and Memory-Light Vision Transformers
2605.15422 (2026-05)
DualKV: Shared-Prompt Flash Attention for Efficient RL Training with Large Rollouts and Long Contexts
round_multiple 3 other papers
1 dated earlier, 2 later
2309.06180 (2023-09)
Efficient Memory Management for Large Language Model Serving with PagedAttention
2604.23798 (2026-04)
ELSA: Exact Linear-Scan Attention for Fast and Memory-Light Vision Transformers
2605.15422 (2026-05)
DualKV: Shared-Prompt Flash Attention for Efficient RL Training with Large Rollouts and Long Contexts
attention_pytorch 2 other papers
1 dated earlier, 1 later
2309.06180 (2023-09)
Efficient Memory Management for Large Language Model Serving with PagedAttention
2604.23798 (2026-04)
ELSA: Exact Linear-Scan Attention for Fast and Memory-Light Vision Transformers
convert_to_cudnn_type 2 other papers
1 dated earlier, 1 later
2309.06180 (2023-09)
Efficient Memory Management for Large Language Model Serving with PagedAttention
2604.23798 (2026-04)
ELSA: Exact Linear-Scan Attention for Fast and Memory-Light Vision Transformers
cudnn_spda_setup 2 other papers
1 dated earlier, 1 later
2309.06180 (2023-09)
Efficient Memory Management for Large Language Model Serving with PagedAttention
2604.23798 (2026-04)
ELSA: Exact Linear-Scan Attention for Fast and Memory-Light Vision Transformers
efficiency 2 other papers
1 dated earlier, 1 later
2309.06180 (2023-09)
Efficient Memory Management for Large Language Model Serving with PagedAttention
2604.23798 (2026-04)
ELSA: Exact Linear-Scan Attention for Fast and Memory-Light Vision Transformers
flops 2 other papers
1 dated earlier, 1 later
2309.06180 (2023-09)
Efficient Memory Management for Large Language Model Serving with PagedAttention
2604.23798 (2026-04)
ELSA: Exact Linear-Scan Attention for Fast and Memory-Light Vision Transformers

Presence, not direction: an identical function body in two papers' repositories does not say who wrote it, which came from which, or that either paper's method rests on it. Matching is exact after normalisation, so a renamed or edited function is not matched. Dates are when papers appeared, not when code was written. Only .py files the harvester lifted are covered; notebooks are not read. A new-style arXiv id YYMM.NNNN dates a paper to the month 20YY-MM; otherwise Paper.year dates it to the year; otherwise it is undated. Dates are when the paper appeared, not when the code was written. Read 2026-11-22.

Repositories linked to this paper

Some links come from the archived Papers with Code dataset (CC BY-SA 4.0): attribution and licence.

Abstract

Attention, as a core layer of the ubiquitous Transformer architecture, is the bottleneck for large language models and long-context applications. FlashAttention elaborated an approach to speed up attention on GPUs through minimizing memory reads/writes. However, it has yet to take advantage of new capabilities present in recent hardware, with FlashAttention-2 achieving only 35% utilization on the H100 GPU. We develop three main techniques to speed up attention on Hopper GPUs: exploiting asynchrony of the Tensor Cores and TMA to (1) overlap overall computation and data movement via warp-specialization and (2) interleave block-wise matmul and softmax operations, and (3) block quantization and incoherent processing that leverages hardware support for FP8 low-precision. We demonstrate that our method, FlashAttention-3, achieves speedup on H100 GPUs by 1.5-2.0$\times$ with FP16 reaching up to 740 TFLOPs/s (75% utilization), and with FP8 reaching close to 1.2 PFLOPs/s. We validate that FP8 FlashAttention-3 achieves 2.6$\times$ lower numerical error than a baseline FP8 attention.

For agents

The same record, over MCP at https://syntology.ai/mcp:

get_harvested_code_for_paper("2407.08608")
get_code_for_paper("2407.08608")
have("2407.08608")

The run record, dated, one paper per request, free:

curl https://syntology.ai/api/ran/2407.08608.json

A badge for a README (the split and the date, never a ratio):

[![Syntology run record](https://syntology.ai/api/ran/2407.08608.svg)](https://syntology.ai/paper/2407.08608)

Connect an agent — have() is free.