Revert "[Perf][GLM-5.2] Blackwell decode optimizations" (#49768)

This commit is contained in:
Woosuk Kwon
2026-07-24 16:04:59 -07:00
committed by GitHub
parent 33c4f3551c
commit 213f681f81
29 changed files with 138 additions and 2139 deletions
@@ -49,11 +49,6 @@ __global__ void __launch_bounds__(512, VLLM_BLOCKS_PER_SM(512))
static_assert(sizeof(PackedVec) == sizeof(Type) * CVT_FP4_ELTS_PER_THREAD,
"Vec size is not matched.");
#if (defined(__CUDA_ARCH__) && (__CUDA_ARCH__ >= 900))
cudaGridDependencySynchronize();
cudaTriggerProgrammaticLaunchCompletion();
#endif
// Precompute SF layout parameter (constant for entire kernel).
int32_t const numKTiles = (outputCols + 63) / 64;
@@ -128,11 +123,6 @@ __global__ void __launch_bounds__(512, VLLM_BLOCKS_PER_SM(512))
static_assert(sizeof(PackedVec) == sizeof(Type) * CVT_FP4_ELTS_PER_THREAD,
"Vec size is not matched.");
#if (defined(__CUDA_ARCH__) && (__CUDA_ARCH__ >= 900))
cudaGridDependencySynchronize();
cudaTriggerProgrammaticLaunchCompletion();
#endif
int32_t const colIdx = blockDim.x * blockIdx.y + threadIdx.x;
int elem_idx = colIdx * CVT_FP4_ELTS_PER_THREAD;
@@ -211,8 +201,6 @@ void scaled_fp4_quant_sm1xxa(torch::stable::Tensor const& output,
const torch::stable::accelerator::DeviceGuard device_guard(
input.get_device_index());
auto stream = get_current_cuda_stream(input.get_device_index());
auto* device_props = get_device_prop();
int const sm_version = device_props->major * 10 + device_props->minor;
int output_sf_n_unpadded = int(output_n / CVT_FP4_SF_VEC_SIZE);
@@ -236,21 +224,10 @@ void scaled_fp4_quant_sm1xxa(torch::stable::Tensor const& output,
input.scalar_type(), "nvfp4_quant_kernel", [&] {
using cuda_type = vllm::CUDATypeConverter<scalar_t>::Type;
auto input_ptr = static_cast<cuda_type const*>(input.data_ptr());
cudaLaunchConfig_t config = {};
config.gridDim = grid;
config.blockDim = block;
config.dynamicSmemBytes = 0;
config.stream = stream;
cudaLaunchAttribute attrs[1];
attrs[0].id = cudaLaunchAttributeProgrammaticStreamSerialization;
attrs[0].val.programmaticStreamSerializationAllowed = 1;
config.numAttrs = (sm_version >= 90) ? 1 : 0;
config.attrs = attrs;
cudaLaunchKernelEx(&config, vllm::cvt_fp16_to_fp4<cuda_type, false>,
m, n, output_n, num_padded_cols, input_ptr,
input_sf_ptr,
reinterpret_cast<uint32_t*>(output_ptr),
reinterpret_cast<uint32_t*>(sf_out));
vllm::cvt_fp16_to_fp4<cuda_type, false><<<grid, block, 0, stream>>>(
m, n, output_n, num_padded_cols, input_ptr, input_sf_ptr,
reinterpret_cast<uint32_t*>(output_ptr),
reinterpret_cast<uint32_t*>(sf_out));
});
} else {
int num_packed_cols = output_n / CVT_FP4_ELTS_PER_THREAD;
@@ -263,21 +240,12 @@ void scaled_fp4_quant_sm1xxa(torch::stable::Tensor const& output,
input.scalar_type(), "nvfp4_quant_kernel", [&] {
using cuda_type = vllm::CUDATypeConverter<scalar_t>::Type;
auto input_ptr = static_cast<cuda_type const*>(input.data_ptr());
cudaLaunchConfig_t config = {};
config.gridDim = grid;
config.blockDim = block;
config.dynamicSmemBytes = 0;
config.stream = stream;
cudaLaunchAttribute attrs[1];
attrs[0].id = cudaLaunchAttributeProgrammaticStreamSerialization;
attrs[0].val.programmaticStreamSerializationAllowed = 1;
config.numAttrs = (sm_version >= 90) ? 1 : 0;
config.attrs = attrs;
cudaLaunchKernelEx(
&config, vllm::cvt_fp16_to_fp4_sf_major<cuda_type, false>, m, n,
output_n, output_sf_n_unpadded, num_packed_cols, input_ptr,
input_sf_ptr, reinterpret_cast<uint32_t*>(output_ptr),
reinterpret_cast<uint32_t*>(sf_out));
vllm::cvt_fp16_to_fp4_sf_major<cuda_type, false>
<<<grid, block, 0, stream>>>(
m, n, output_n, output_sf_n_unpadded, num_packed_cols,
input_ptr, input_sf_ptr,
reinterpret_cast<uint32_t*>(output_ptr),
reinterpret_cast<uint32_t*>(sf_out));
});
}
}