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 17, 01:45 UTC
[Bugfix] Fix marlin kernel crash on H100 (#4218)
This PR addresses the Marlin kernel H100 crash that was reported here: neuralmagic#187. The reason for the crash was the inline PTX assembly that introduced the async_copy with streaming behavior. The solution is to use the more standard PTX for async_copy (without the fractional L2 policy for "evict_first"). There is no performance difference between standard async_copy PTX and the previous one.
alexm-nm committed
Apr 24, 2024 at 13:35 UTC
aae08249acca69060d0a8220cab920e00520932c
1 file changed
+8
-15
csrc/quantization/marlin/marlin_cuda_kernel.cu
+8
-15
@@ -67,20 +67,13 @@ __device__ inline void cp_async4_pred(void *smem_ptr, const void *glob_ptr,
67
"r"(smem), "l"(glob_ptr), "n"(BYTES));
68
}
69
70
-// Asynchronous global->shared copy with a cache hint indicating that the values
71
-// may be evicted immediately; used for quantized weights B, which are only
72
-// accessed precisely once and should thus not pollute the L2 cache which we
73
-// need for inputs A and outputs C.
74
-__device__ inline void cp_async4_stream(void *smem_ptr, const void *glob_ptr) {
70
+// Asynchronous global->shared copy
71
+__device__ inline void cp_async4(void *smem_ptr, const void *glob_ptr) {
72
const int BYTES = 16;
73
uint32_t smem = static_cast<uint32_t>(__cvta_generic_to_shared(smem_ptr));
77
- asm volatile(
78
- "{\n"
79
- " .reg .b64 p;\n"
80
- " createpolicy.fractional.L2::evict_first.b64 p, 1.0;"
81
- " cp.async.cg.shared.global.L2::cache_hint [%0], [%1], %2, p;\n"
82
- "}\n" ::"r"(smem),
83
- "l"(glob_ptr), "n"(BYTES));
74
+ asm volatile("{\n"
75
+ " cp.async.cg.shared.global [%0], [%1], %2;\n"
76
+ "}\n" :: "r"(smem), "l"(glob_ptr), "n"(BYTES));
77
}
78
79
// Async copy fence.
@@ -448,14 +441,14 @@ Marlin(const int4 *__restrict__ A, // fp16 input matrix of shape mxk
441
int4 *sh_b_stage = sh_b + b_sh_stage * pipe;
442
#pragma unroll
443
for (int i = 0; i < b_sh_wr_iters; i++) {
451
- cp_async4_stream(&sh_b_stage[b_sh_wr_delta * i + b_sh_wr], B_ptr[i]);
444
+ cp_async4(&sh_b_stage[b_sh_wr_delta * i + b_sh_wr], B_ptr[i]);
445
B_ptr[i] += b_gl_rd_delta_o;
446
}
447
// Only fetch scales if this tile starts a new group
448
if (group_blocks != -1 && pipe % (group_blocks / thread_k_blocks) == 0) {
449
int4 *sh_s_stage = sh_s + s_sh_stage * pipe;
450
if (s_sh_wr_pred)
458
- cp_async4_stream(&sh_s_stage[s_sh_wr], &s[s_gl_rd]);
451
+ cp_async4(&sh_s_stage[s_sh_wr], &s[s_gl_rd]);
452
s_gl_rd += s_gl_rd_delta;
453
}
454
}
@@ -750,7 +743,7 @@ Marlin(const int4 *__restrict__ A, // fp16 input matrix of shape mxk
743
// write-out
744
if (group_blocks == -1 && last) {
745
if (s_sh_wr_pred)
753
- cp_async4_stream(&sh_s[s_sh_wr], &s[s_gl_rd]);
746
+ cp_async4(&sh_s[s_sh_wr], &s[s_gl_rd]);
747
cp_async_fence();
748
}
749
thread_block_reduce();