v4.3 update. (#2709)

* v4.3 update.

* Update the cute_dsl_api changelog's doc link

* Update version to 4.3.0

* Update the example link

* Update doc to encourage user to install DSL from requirements.txt

---------

Co-authored-by: Larry Wu <larwu@nvidia.com>
This commit is contained in:
Junkai-Wu
2025-10-21 14:26:30 -04:00
committed by GitHub
co-authored by Larry Wu
parent e6e2cc29f5
commit b1d6e2c9b3
244 changed files with 59272 additions and 10455 deletions
@@ -329,6 +329,20 @@ void fmha_bwd_reference_dQ(
dim3 grid(size<0>(mDQ), size<2>(mDQ), 1);
dim3 block(256);
int shared_mem = size<0>(mK) * sizeof(typename TensorDQ::value_type);
cudaError_t result;
if (shared_mem >= (48 << 10)) {
result = cudaFuncSetAttribute(
&fmha_bwd_reference_dQ_kernel<ProblemShape, TensorQ, TensorK, TensorV,
TensorO, TensorLSE, TensorDO, TensorDQ, Fusion>,
cudaFuncAttributeMaxDynamicSharedMemorySize,
shared_mem);
if (cudaSuccess != result) {
cudaGetLastError(); // Clear the error state
throw std::runtime_error("Failed to allocate " +
std::to_string(shared_mem >> 10) + " KB dynamic smem for dQ tensor in ref. check - " +
"please try reducing seq_len or skipping ref. check");
}
}
fmha_bwd_reference_dQ_kernel<<<grid, block, shared_mem>>>(problem_shape, mQ, mK, mV, mO, mLSE, mDO, mDQ, fusion);
}
@@ -356,6 +370,20 @@ void fmha_bwd_reference_dK(
dim3 grid(K, H_K * B, 1);
dim3 block(std::max(D, 256));
int shared_mem = size<0>(mDO) * sizeof(typename TensorDK::value_type);
cudaError_t result;
if (shared_mem >= (48 << 10)) {
result = cudaFuncSetAttribute(
&fmha_bwd_reference_dK_kernel<ProblemShape, TensorQ, TensorK, TensorV,
TensorO, TensorLSE, TensorDO, TensorDK, Fusion>,
cudaFuncAttributeMaxDynamicSharedMemorySize,
shared_mem);
if (cudaSuccess != result) {
cudaGetLastError(); // Clear the error state
throw std::runtime_error("Failed to allocate " +
std::to_string(shared_mem >> 10) + " KB dynamic smem for dO tensor in ref. check - " +
"please try reducing seq_len or skipping ref. check");
}
}
fmha_bwd_reference_dK_kernel<<<grid, block, shared_mem>>>(problem_shape, mQ, mK, mV, mO, mLSE, mDO, mDK, fusion);
}
@@ -383,6 +411,20 @@ void fmha_bwd_reference_dV(
dim3 grid(K, H_K * B, 1);
dim3 block(std::max(D_VO, 256));
int shared_mem = size<0>(mDO) * sizeof(typename TensorDV::value_type);
cudaError_t result;
if (shared_mem >= (48 << 10)) {
result = cudaFuncSetAttribute(
&fmha_bwd_reference_dV_kernel<ProblemShape, TensorQ, TensorK, TensorV,
TensorO, TensorLSE, TensorDO, TensorDV, Fusion>,
cudaFuncAttributeMaxDynamicSharedMemorySize,
shared_mem);
if (cudaSuccess != result) {
cudaGetLastError(); // Clear the error state
throw std::runtime_error("Failed to allocate " +
std::to_string(shared_mem >> 10) + " KB dynamic smem for dO tensor in ref. check - " +
"please try reducing seq_len or skipping ref. check");
}
}
fmha_bwd_reference_dV_kernel<<<grid, block, shared_mem>>>(problem_shape, mQ, mK, mV, mO, mLSE, mDO, mDV, fusion);
}
@@ -189,6 +189,19 @@ void fmha_reference(
dim3 grid(size<0>(mO), size<2>(mO), 1);
dim3 block(256);
int shared_mem = size<0>(mK) * int(sizeof(typename TensorLSE::value_type));
cudaError_t result;
if (shared_mem >= (48 << 10)) {
result = cudaFuncSetAttribute(
&fmha_reference_kernel<ProblemShapeIn, TensorQ, TensorK, TensorV, TensorO, TensorLSE, Mask>,
cudaFuncAttributeMaxDynamicSharedMemorySize,
shared_mem);
if (cudaSuccess != result) {
cudaGetLastError(); // Clear the error state
throw std::runtime_error("Failed to allocate " +
std::to_string(shared_mem >> 10) + " KB dynamic smem for S/P tensor in ref. check - " +
"please try reducing seq_len or skipping ref. check");
}
}
fmha_reference_kernel<<<grid, block, shared_mem>>>(problem_shape_in, mQ, mK, mV, mO, mLSE, mask);
}
@@ -185,9 +185,11 @@ void fmha_mla_reference(
cudaFuncAttributeMaxDynamicSharedMemorySize,
shared_mem);
if (cudaSuccess != result) {
result = cudaGetLastError(); // to clear the error bit
throw std::runtime_error("couldn't perform smem optin");
}
cudaGetLastError(); // Clear the error state
throw std::runtime_error("Failed to allocate " +
std::to_string(shared_mem >> 10) + " KB dynamic smem for S/P tensor in ref. check - " +
"please try reducing seq_len or skipping ref. check");
}
}
fmha_mla_reference_kernel<<<grid, block, shared_mem>>>(
problem_shape, mSeq, mPT, mQL, mQR, mCL, mKR, mO, mLSE, scale);