@tduncan / Nvidia-Optimization-Stack-VLLM Mirror · read-only / commits / cbbc904470
Fork 0

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 16, 19:30 UTC

[Kernel] Squash a few more warnings (#6914)

Tyler Michael Smith committed Jul 30, 2024 at 13:50 UTC cbbc904470668b9420e71595edeef76d673a2d59
5 files changed +8 -5
csrc/attention/attention_kernels.cu
+2 -2
@@ -706,7 +706,7 @@ void paged_attention_v1_launcher(
706 int kv_block_stride = key_cache.stride(0);
707 int kv_head_stride = key_cache.stride(1);
708
709 - int thread_group_size = MAX(WARP_SIZE / BLOCK_SIZE, 1);
709 + [[maybe_unused]] int thread_group_size = MAX(WARP_SIZE / BLOCK_SIZE, 1);
710 assert(head_size % thread_group_size == 0);
711
712 // NOTE: alibi_slopes is optional.
@@ -865,7 +865,7 @@ void paged_attention_v2_launcher(
865 int kv_block_stride = key_cache.stride(0);
866 int kv_head_stride = key_cache.stride(1);
867
868 - int thread_group_size = MAX(WARP_SIZE / BLOCK_SIZE, 1);
868 + [[maybe_unused]] int thread_group_size = MAX(WARP_SIZE / BLOCK_SIZE, 1);
869 assert(head_size % thread_group_size == 0);
870
871 // NOTE: alibi_slopes is optional.
csrc/quantization/aqlm/gemm_kernels.cu
-2
@@ -273,8 +273,6 @@ __global__ void Code2x8Dequant(
273 }
274 __syncthreads();
275
276 - float res = 0;
277 -
276 int iters = (prob_k / 8 - 1) / (8 * 32) + 1;
277 while (iters--) {
278 if (pred && a_gl_rd < a_gl_end) {
csrc/quantization/fp8/amd/quant_utils.cuh
+2
@@ -526,6 +526,7 @@ __inline__ __device__ Tout convert(const Tin& x) {
526 }
527 #endif
528 assert(false);
529 + return {}; // Squash missing return statement warning
530 }
531
532 template <typename Tout, typename Tin, Fp8KVCacheDataType kv_dt>
@@ -536,6 +537,7 @@ __inline__ __device__ Tout scaled_convert(const Tin& x, const float scale) {
537 }
538 #endif
539 assert(false);
540 + return {}; // Squash missing return statement warning
541 }
542
543 // The following macro is used to dispatch the conversion function based on
csrc/quantization/fp8/nvidia/quant_utils.cuh
+2
@@ -508,6 +508,7 @@ __inline__ __device__ Tout convert(const Tin& x) {
508 }
509 #endif
510 assert(false);
511 + return {}; // Squash missing return statement warning
512 }
513
514 template <typename Tout, typename Tin, Fp8KVCacheDataType kv_dt>
@@ -520,6 +521,7 @@ __inline__ __device__ Tout scaled_convert(const Tin& x, const float scale) {
521 }
522 #endif
523 assert(false);
524 + return {}; // Squash missing return statement warning
525 }
526
527 // The following macro is used to dispatch the conversion function based on
csrc/quantization/squeezellm/quant_cuda_kernel.cu
+2 -1
@@ -203,7 +203,8 @@ void squeezellm_gemm(torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
203 #endif
204 mat.data_ptr<int>(),
205 #ifndef USE_ROCM
206 - (half2*)mul.data<at::Half>(), (__half*)lookup_table.data_ptr<at::Half>(),
206 + (half2*)mul.data_ptr<at::Half>(),
207 + (__half*)lookup_table.data_ptr<at::Half>(),
208 #else
209 (float2*)mul.data_ptr<float>(),
210 (__half*)lookup_table.data_ptr<at::Half>(),