squall/flash-attention
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

[Docs] Clearer error message for bwd d > 64, bump to v1.0.4

Browse Source
This commit is contained in:
Tri Dao 2023-04-26 09:19:48 -07:00
parent fbbb107848
commit ad113948a6
3 changed files with 5 additions and 5 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
csrc/flash_attn/fmha_api.cpp
View File

@ -413,8 +413,8 @@ mha_bwd(const at::Tensor &dout, // total_q x num_heads, x head_size
const int total_k = k.size(TOTAL_DIM);
TORCH_CHECK(batch_size > 0);
TORCH_CHECK((head_size % 8 == 0) && (head_size <= 128));
if (head_size > 64) { // TODO: eventually we should support SM86 and SM70 with d=128 as well
TORCH_CHECK(is_sm80 || is_sm90);
if (head_size > 64) {
TORCH_CHECK(is_sm80 || is_sm90, "FlashAttention backward for head dim > 64 requires A100 or H100 GPUs as the implementation needs a large amount of shared memory.");
}
CHECK_SHAPE(q, total_q, num_heads, head_size);

2
setup.py
View File

@ -162,7 +162,7 @@ ext_modules.append(
setup(
name="flash_attn",
version="1.0.3.post0",
version="1.0.4",
packages=find_packages(
exclude=("build", "csrc", "include", "tests", "dist", "docs", "benchmarks", "flash_attn.egg-info",)
),

4
training/Dockerfile
View File

@ -85,11 +85,11 @@ RUN pip install transformers==4.25.1 datasets==2.8.0 pytorch-lightning==1.8.6 tr
RUN pip install git+https://github.com/mlcommons/logging.git@2.1.0
# Install FlashAttention
RUN pip install flash-attn==1.0.3.post0
RUN pip install flash-attn==1.0.4
# Install CUDA extensions for cross-entropy, fused dense, layer norm
RUN git clone https://github.com/HazyResearch/flash-attention \
&& cd flash-attention && git checkout v1.0.3.post0 \
&& cd flash-attention && git checkout v1.0.4 \
&& cd csrc/fused_softmax && pip install . && cd ../../ \
&& cd csrc/rotary && pip install . && cd ../../ \
&& cd csrc/xentropy && pip install . && cd ../../ \