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 11, 00:00 UTC
[DeepSeek v3.2] Remove unnecessary syncwarps (#31047)
Signed-off-by: Matthew Bonanni <mbonanni@redhat.com>
Matthew Bonanni committed
Dec 24, 2025 at 00:33 UTC
369f47aa0f1c8ab8e24a13c1dde7d1626f3b940e
1 file changed
+1
-6
csrc/cache_kernels.cu
+1
-6
@@ -451,9 +451,6 @@ __global__ void indexer_k_quant_and_cache_kernel(
451
for (int i = 0; i < VEC_SIZE; i++) {
452
amax = fmaxf(amax, fabsf(float(k_val_ptr[i])));
453
}
454
-#ifndef USE_ROCM
455
- __syncwarp();
456
-#endif
454
455
// Reduced amax
456
for (int mask = 16; mask > 0; mask /= 2) {
@@ -463,9 +460,7 @@ __global__ void indexer_k_quant_and_cache_kernel(
460
amax = fmaxf(amax, __shfl_xor_sync(unsigned(-1), amax, mask));
461
#endif
462
}
466
-#ifndef USE_ROCM
467
- __syncwarp();
468
-#endif
463
+
464
#if defined(__gfx942__)
465
float scale = fmaxf(amax, 1e-4) / 224.0f;
466
#else