A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 15, 04:45 UTC
[Bugfix] Fix some narrowing conversion warnings (#20141)
Signed-off-by: Tyler Michael Smith <tyler@neuralmagic.com>
Tyler Michael Smith committed
Jun 27, 2025 at 12:01 UTC
e8c3bd2cd164786e764f7e3436bbaa6cd00ae64a
6 files changed
+8
-14
csrc/attention/mla/cutlass_mla_kernels.cu
+1
-1
@@ -207,7 +207,7 @@ void cutlass_mla_decode_sm100a(torch::Tensor const& out,
207
"page_table must be a 32-bit integer tensor");
208
209
auto in_dtype = q_nope.dtype();
210
- at::cuda::CUDAGuard device_guard{(char)q_nope.get_device()};
210
+ const at::cuda::OptionalCUDAGuard device_guard(device_of(q_nope));
211
const cudaStream_t stream =
212
at::cuda::getCurrentCUDAStream(q_nope.get_device());
213
if (in_dtype == at::ScalarType::Half) {
csrc/mamba/causal_conv1d/causal_conv1d.cu
+2
-6
@@ -185,9 +185,7 @@ void causal_conv1d_fwd(const at::Tensor &x, const at::Tensor &weight,
185
params.conv_states_ptr = nullptr;
186
}
187
188
- // Otherwise the kernel will be launched from cuda:0 device
189
- // Cast to char to avoid compiler warning about narrowing
190
- at::cuda::CUDAGuard device_guard{(char)x.get_device()};
188
+ const at::cuda::OptionalCUDAGuard device_guard(device_of(x));
189
auto stream = at::cuda::getCurrentCUDAStream().stream();
190
DISPATCH_WTYPE_ITYPE_FLOAT_AND_HALF_AND_BF16(x.scalar_type(), "causal_conv1d_fwd", [&] {
191
causal_conv1d_fwd_cuda<input_t, weight_t>(params, stream);
@@ -278,9 +276,7 @@ void causal_conv1d_update(const at::Tensor &x,
276
params.conv_state_indices_ptr = nullptr;
277
}
278
281
- // Otherwise the kernel will be launched from cuda:0 device
282
- // Cast to char to avoid compiler warning about narrowing
283
- at::cuda::CUDAGuard device_guard{(char)x.get_device()};
279
+ const at::cuda::OptionalCUDAGuard device_guard(device_of(x));
280
auto stream = at::cuda::getCurrentCUDAStream().stream();
281
DISPATCH_WTYPE_ITYPE_FLOAT_AND_HALF_AND_BF16(x.scalar_type(), "causal_conv1d_update", [&] {
282
causal_conv1d_update_cuda<input_t, weight_t>(params, stream);
csrc/mamba/mamba_ssm/selective_scan_fwd.cu
+1
-3
@@ -647,9 +647,7 @@ void selective_scan_fwd(const torch::Tensor &u, const torch::Tensor &delta,
647
);
648
649
650
- // Otherwise the kernel will be launched from cuda:0 device
651
- // Cast to char to avoid compiler warning about narrowing
652
- at::cuda::CUDAGuard device_guard{(char)u.get_device()};
650
+ const at::cuda::OptionalCUDAGuard device_guard(device_of(u));
651
auto stream = at::cuda::getCurrentCUDAStream().stream();
652
DISPATCH_WTYPE_ITYPE_FLOAT_AND_HALF_AND_BF16(u.scalar_type(), "selective_scan_fwd", [&] {
653
selective_scan_fwd_cuda<input_t, weight_t>(params, stream);
csrc/quantization/fp4/nvfp4_experts_quant.cu
+2
-2
@@ -561,7 +561,7 @@ void scaled_fp4_experts_quant_sm100a(
561
TORCH_CHECK(output_scale.size(1) * 4 == padded_k);
562
563
auto in_dtype = input.dtype();
564
- at::cuda::CUDAGuard device_guard{(char)input.get_device()};
564
+ const at::cuda::OptionalCUDAGuard device_guard(device_of(input));
565
const cudaStream_t stream =
566
at::cuda::getCurrentCUDAStream(input.get_device());
567
if (in_dtype == at::ScalarType::Half) {
@@ -579,4 +579,4 @@ void scaled_fp4_experts_quant_sm100a(
579
} else {
580
TORCH_CHECK(false, "Expected input data type to be half or bfloat16");
581
}
582
-}
\ No newline at end of file
582
+}
csrc/quantization/fp4/nvfp4_quant_kernels.cu
+1
-1
@@ -347,7 +347,7 @@ void scaled_fp4_quant_sm100a(torch::Tensor const& output,
347
auto input_sf_ptr = static_cast<float const*>(input_sf.data_ptr());
348
auto sf_out = static_cast<int32_t*>(output_sf.data_ptr());
349
auto output_ptr = static_cast<int64_t*>(output.data_ptr());
350
- at::cuda::CUDAGuard device_guard{(char)input.get_device()};
350
+ const at::cuda::OptionalCUDAGuard device_guard(device_of(input));
351
auto stream = at::cuda::getCurrentCUDAStream(input.get_device());
352
353
// We don't support e8m0 scales at this moment.
csrc/quantization/fp4/nvfp4_scaled_mm_kernels.cu
+1
-1
@@ -267,7 +267,7 @@ void cutlass_scaled_fp4_mm_sm100a(torch::Tensor& D, torch::Tensor const& A,
267
B_sf.sizes()[1], ")");
268
269
auto out_dtype = D.dtype();
270
- at::cuda::CUDAGuard device_guard{(char)A.get_device()};
270
+ const at::cuda::OptionalCUDAGuard device_guard(device_of(A));
271
const cudaStream_t stream = at::cuda::getCurrentCUDAStream(A.get_device());
272
273
if (out_dtype == at::ScalarType::Half) {