Commits

Branch selector

 main

User selector

All users

Datepicker

All time

Commit History

Commits on Oct 6, 2022

• fix failing tests facebookresearch#464 due to windows distributed using gloo not nccl (facebookresearch#465)

  Show description for 6e96003
  Thomas-MMJ

  authored
  6e96003

• conda fix

  Show description for 4bc9ffb
  bottler

  committed
  4bc9ffb

• Fix compiler errors on Windows (facebookresearch#460)

  Show description for 1583589
  yocabon

  authored
  1583589

• info: Add 'python -m xformers.info' to debug setup

  Show description for d1e4611
  

  danthe3rd

  committed
  d1e4611

• chunk3: Add custom operator to avoid torch.cat in BW

  Show description for 6b62c6b
  

  danthe3rd

  committed
  6b62c6b

• bwpackedgrad: Pack QKV gradients as well

  Show description for e92935c
  

  danthe3rd

  committed
  e92935c

• bwpacked: Support non-contiguous Q/K/V for backward

  Show description for d2083ed
  

  danthe3rd

  committed
  d2083ed

• bwinline: Forceinline all functions of the bw

  Show description for 1940d0d
  

  danthe3rd

  committed
  1940d0d

• v100smem: Write in 32bits blocks to smem on V100

  Show description for 85d4ba9
  

  danthe3rd

  committed
  85d4ba9

Commits on Oct 3, 2022

• p100nvcc2: Better workaround for compiler bug on P100

  Show description for b5a1cb9
  

  danthe3rd

  committed
  b5a1cb9

• packed: Support (some) non-contiguous Q,K,V

  Show description for 7e4c02c
  

  danthe3rd

  committed
  7e4c02c

• log2ef: Multiply by M_LOG2E in float (rather than double)

  Show description for b02f013
  

  danthe3rd

  committed
  b02f013

• v100reduc: Implement warp reduction on V100 for same line [5-10% improv]

  Show description for 6c521e1
  

  danthe3rd

  committed
  6c521e1

• fw_forceinline: Apply forceinline for all device functions

  Show description for 1ab2221
  

  danthe3rd

  committed
  1ab2221

• regfix: Fix regression with cu seqlen PR

  Show description for 4025c37
  

  danthe3rd

  committed
  4025c37

• ci: Enable CI on multiple GPUs

  Show description for 8e77581
  

  danthe3rd

  committed
  8e77581

Commits on Sep 30, 2022

• [nit] Remove the "Error" wording when no triton is installed (facebookresearch#450)

  Show description for ae6afaa
  

  blefaudeux

  and

  Benjamin Lefaudeux

  authored
  ae6afaa

Commits on Sep 29, 2022

• building conda packages

  Show description for 9063515
  bottler

  committed
  9063515

• v100nvcc: work around a compiler bug in V100

  Show description for 132a148
  

  danthe3rd

  committed
  132a148

• k_kv: support value.shape[-1] != query.shape[-1]

  Show description for 6e690df
  

  danthe3rd

  committed
  6e690df

• seqlen: Support variable sequence lengths

  Show description for fa3127a
  

  danthe3rd

  committed
  fa3127a

• noqs: Advance pointers to block position

  Show description for d32f24f
  

  danthe3rd

  committed
  d32f24f

• cleanup: Cleanup forward/backward dispatch

  Show description for 987b98b
  

  danthe3rd

  committed
  987b98b

• p100nvcc: Workaround a compiler bug on P100

  Show description for 06f801d
  

  danthe3rd

  committed
  06f801d

• prologuemma: Prefetch tensors before mmuls

  Show description for 7f9d25f
  

  danthe3rd

  committed
  7f9d25f

• causalfix: Fix causal after iteration inversion

  Show description for 0c7697c
  

  danthe3rd

  committed
  0c7697c

• mmafix: Fix matmul out-of-bounds for small dimensions with preloading

  Show description for 41d31e7
  

  danthe3rd

  committed
  41d31e7

• custommma: Add custom MMA for more optimizations

  Show description for a2d1b82
  

  danthe3rd

  committed
  a2d1b82

• bf16: Add support in fw/bw

  Show description for 7825f7c
  

  danthe3rd

  committed
  7825f7c

• outputf16: Change forward's output to match input dtype

  Show description for 18c875e
  

  danthe3rd

  committed
  18c875e

• bw_preproc: Compute delta inside kernels for f16

  Show description for fec6b2a
  

  danthe3rd

  committed
  fec6b2a

• noinit: Don't initialize backward's gradients to zero

  Show description for 8071784
  

  danthe3rd

  committed
  8071784

• Preload operands for GradQ/GradV on A100

  Show description for 1bf346f
  

  danthe3rd

  committed
  1bf346f

• Pipeline loads in epilogue (~5% speedup)

  Show description for 3c7dc91
  

  danthe3rd

  committed
  3c7dc91

• Prefetch output tensor in epilogue

  Show description for c843654
  

  danthe3rd

  committed
  c843654

Pagination

PreviousNext