Merge branch 'HazyResearch:main' into enable_cuda_graph_capture

README.md

@@ -38,7 +38,7 @@ and experiment with. The notations in the Triton implementation are also closer
 to what's used in our paper.
 
 
-## Beta release (0.2).
+## Installation and features
 
 Requirements:
 - CUDA 11.4 and above.

csrc/fused_dense_lib/fused_dense_cuda.cu

@@ -122,7 +122,9 @@ int gemm_bias_act_lt(
     reinterpret_cast<cublasLtHandle_t>(at::cuda::getCurrentCUDABlasHandle());
   // See https://github.com/pytorch/pytorch/issues/73328 for reasoning behind
   // setting this to 1M.
-  size_t workspaceSize = 1024 * 1024;
+  // However, Apex sets it to 4M and TransformerEngine sets to 32M for Hopper and 4M for other GPUs
+  // https://github.com/NVIDIA/TransformerEngine/blob/a0f0065498bbcfc1da78cf9e8b166f5381613fbc/transformer_engine/pytorch/module.py#L91
+  size_t workspaceSize = 1024 * 1024 * (at::cuda::getCurrentDeviceProperties()->major >= 9 ? 32 : 4);
   void* workspace = at::empty(
     {static_cast<int64_t>(workspaceSize)},
     at::device({at::kCUDA, at::cuda::current_device()}).dtype(at::kByte)).data_ptr();
@@ -296,7 +298,8 @@ int gemm_bgradb_lt(
     reinterpret_cast<cublasLtHandle_t>(at::cuda::getCurrentCUDABlasHandle());
   // See https://github.com/pytorch/pytorch/issues/73328 for reasoning behind
   // setting this to 1M.
-  size_t workspaceSize = 1024 * 1024;
+  // However, Apex sets it to 4M and TransformerEngine sets to 32M for Hopper and 4M for other GPUs
+  size_t workspaceSize = 1024 * 1024 * (at::cuda::getCurrentDeviceProperties()->major >= 9 ? 32 : 4);
   void* workspace = at::empty(
     {static_cast<int64_t>(workspaceSize)},
     at::device({at::kCUDA, at::cuda::current_device()}).dtype(at::kByte)).data_ptr();
@@ -449,7 +452,8 @@ int gemm_dact_bgradb_lt(
     reinterpret_cast<cublasLtHandle_t>(at::cuda::getCurrentCUDABlasHandle());
   // See https://github.com/pytorch/pytorch/issues/73328 for reasoning behind
   // setting this to 1M.
-  size_t workspaceSize = 1024 * 1024;
+  // However, Apex sets it to 4M and TransformerEngine sets to 32M for Hopper and 4M for other GPUs
+  size_t workspaceSize = 1024 * 1024 * (at::cuda::getCurrentDeviceProperties()->major >= 9 ? 32 : 4);
   void* workspace = at::empty(
     {static_cast<int64_t>(workspaceSize)},
     at::device({at::kCUDA, at::cuda::current_device()}).dtype(at::kByte)).data_ptr();

flash_attn/modules/mlp.py

@@ -17,7 +17,7 @@ class Mlp(nn.Module):
         factory_kwargs = {'device': device, 'dtype': dtype}
         super().__init__()
         out_features = out_features or in_features
-        hidden_features = hidden_features or in_features
+        hidden_features = hidden_features or in_features * 4
         self.return_residual = return_residual
         self.fc1 = nn.Linear(in_features, hidden_features, **factory_kwargs)
         self.activation = activation

setup.py

@@ -162,7 +162,7 @@ ext_modules.append(
 
 setup(
     name="flash_attn",
-    version="0.2.8",
+    version="1.0.1",
     packages=find_packages(
         exclude=("build", "csrc", "include", "tests", "dist", "docs", "benchmarks", "flash_attn.egg-info",)
     ),

training/Dockerfile

@@ -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==0.2.8
+RUN pip install flash-attn==1.0.1
 
 # Install CUDA extensions for cross-entropy, fused dense, layer norm
 RUN git clone https://github.com/HazyResearch/flash-attention \
-    && cd flash-attention && git checkout v0.2.8 \
+    && cd flash-attention && git checkout v1.0.1 \
     && cd csrc/fused_softmax && pip install . && cd ../../ \
     && cd csrc/rotary && pip install . && cd ../../ \
     && cd csrc/xentropy && pip install . && cd ../../ \