tqcq/llama.cpp
0
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-08-09 18:40:20 -04:00
Code Issues Packages Projects Releases Wiki Activity

ggml : reduce hash table reset cost (#8698)

Browse Source 
* ggml : reduce hash table reset cost

* fix unreachable code warnings after GGML_ASSERT(false)

* GGML_ASSERT(false) -> GGML_ABORT("fatal error")

* GGML_ABORT use format string
This commit is contained in:
slaren
2024-07-27 04:41:55 +02:00
committed by GitHub
parent 01245f5b16
commit 2b1f616b20
46 changed files with 851 additions and 754 deletions
Download Patch File Download Diff File Expand all files Collapse all files

42
ggml/src/ggml-metal.m
View File

@@ -869,7 +869,7 @@ static enum ggml_status ggml_metal_graph_compute(
NSError * error = nil;
if (![[MTLCaptureManager sharedCaptureManager] startCaptureWithDescriptor:descriptor error:&error]) {
GGML_METAL_LOG_ERROR("%s: error: unable to start capture '%s'\n", __func__, [[error localizedDescription] UTF8String]);
GGML_ASSERT(!"capture failed");
GGML_ABORT("capture failed");
}
}
@@ -931,7 +931,7 @@ static enum ggml_status ggml_metal_graph_compute(
if (!ggml_metal_supports_op(ctx, dst)) {
GGML_METAL_LOG_ERROR("%s: error: unsupported op '%s'\n", __func__, ggml_op_desc(dst));
GGML_ASSERT(!"unsupported op");
GGML_ABORT("unsupported op");
}
if (should_capture) {
@@ -1068,7 +1068,7 @@ static enum ggml_status ggml_metal_graph_compute(
case GGML_OP_ADD: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ADD_ROW].pipeline; break;
case GGML_OP_MUL: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_ROW].pipeline; break;
case GGML_OP_DIV: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_DIV_ROW].pipeline; break;
default: GGML_ASSERT(false);
default: GGML_ABORT("fatal error");
}
bcast_row = true;
@@ -1077,7 +1077,7 @@ static enum ggml_status ggml_metal_graph_compute(
case GGML_OP_ADD: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ADD].pipeline; break;
case GGML_OP_MUL: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL].pipeline; break;
case GGML_OP_DIV: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_DIV].pipeline; break;
default: GGML_ASSERT(false);
default: GGML_ABORT("fatal error");
}
}
@@ -1131,7 +1131,7 @@ static enum ggml_status ggml_metal_graph_compute(
case GGML_TYPE_F16: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_REPEAT_F16].pipeline; break;
case GGML_TYPE_I32: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_REPEAT_I32].pipeline; break;
case GGML_TYPE_I16: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_REPEAT_I16].pipeline; break;
default: GGML_ASSERT(false);
default: GGML_ABORT("fatal error");
}
[encoder setComputePipelineState:pipeline];
@@ -1387,7 +1387,7 @@ static enum ggml_status ggml_metal_graph_compute(
default:
{
GGML_METAL_LOG_WARN("%s: node %3d, op = %8s not implemented\n", __func__, i, ggml_op_name(dst->op));
GGML_ASSERT(false);
GGML_ABORT("fatal error");
}
} break;
case GGML_OP_SQR:
@@ -1609,7 +1609,7 @@ static enum ggml_status ggml_metal_graph_compute(
case GGML_TYPE_IQ1_M: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ1_M_F32 ].pipeline; break;
case GGML_TYPE_IQ4_NL: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ4_NL_F32 ].pipeline; break;
case GGML_TYPE_IQ4_XS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ4_XS_F32 ].pipeline; break;
default: GGML_ASSERT(false && "MUL MAT-MAT not implemented");
default: GGML_ABORT("MUL MAT-MAT not implemented");
}
[encoder setComputePipelineState:pipeline];
@@ -1782,7 +1782,7 @@ static enum ggml_status ggml_metal_graph_compute(
default:
{
GGML_METAL_LOG_ERROR("Asserting on type %d\n", (int)src0t);
GGML_ASSERT(false && "not implemented");
GGML_ABORT("not implemented");
}
};
@@ -1911,7 +1911,7 @@ static enum ggml_status ggml_metal_graph_compute(
case GGML_TYPE_IQ1_M: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ1_M_F32 ].pipeline; break;
case GGML_TYPE_IQ4_NL: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ4_NL_F32 ].pipeline; break;
case GGML_TYPE_IQ4_XS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ4_XS_F32 ].pipeline; break;
default: GGML_ASSERT(false && "MUL_MAT_ID not implemented");
default: GGML_ABORT("MUL_MAT_ID not implemented");
}
[encoder setComputePipelineState:pipeline];
@@ -2078,7 +2078,7 @@ static enum ggml_status ggml_metal_graph_compute(
default:
{
GGML_METAL_LOG_ERROR("Asserting on type %d\n", (int)src2t);
GGML_ASSERT(false && "not implemented");
GGML_ABORT("not implemented");
}
};
@@ -2178,7 +2178,7 @@ static enum ggml_status ggml_metal_graph_compute(
case GGML_TYPE_IQ4_NL: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ4_NL ].pipeline; break;
case GGML_TYPE_IQ4_XS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ4_XS ].pipeline; break;
case GGML_TYPE_I32: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_I32 ].pipeline; break;
default: GGML_ASSERT(false && "not implemented");
default: GGML_ABORT("not implemented");
}
[encoder setComputePipelineState:pipeline];
@@ -2316,13 +2316,13 @@ static enum ggml_status ggml_metal_graph_compute(
switch (src0->type) {
case GGML_TYPE_F32: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ROPE_NORM_F32].pipeline; break;
case GGML_TYPE_F16: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ROPE_NORM_F16].pipeline; break;
default: GGML_ASSERT(false);
default: GGML_ABORT("fatal error");
};
} else {
switch (src0->type) {
case GGML_TYPE_F32: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ROPE_NEOX_F32].pipeline; break;
case GGML_TYPE_F16: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ROPE_NEOX_F16].pipeline; break;
default: GGML_ASSERT(false);
default: GGML_ABORT("fatal error");
};
}
@@ -2399,7 +2399,7 @@ static enum ggml_status ggml_metal_graph_compute(
switch (dst->type) {
case GGML_TYPE_F32: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_IM2COL_F32].pipeline; break;
case GGML_TYPE_F16: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_IM2COL_F16].pipeline; break;
default: GGML_ASSERT(false);
default: GGML_ABORT("fatal error");
};
[encoder setComputePipelineState:pipeline];
@@ -2556,7 +2556,7 @@ static enum ggml_status ggml_metal_graph_compute(
switch (order) {
case GGML_SORT_ORDER_ASC: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ARGSORT_F32_I32_ASC].pipeline; break;
case GGML_SORT_ORDER_DESC: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ARGSORT_F32_I32_DESC].pipeline; break;
default: GGML_ASSERT(false);
default: GGML_ABORT("fatal error");
};
[encoder setComputePipelineState:pipeline];
@@ -2645,7 +2645,7 @@ static enum ggml_status ggml_metal_graph_compute(
{
GGML_METAL_LOG_ERROR("unsupported size: %lld\n", ne00);
GGML_METAL_LOG_ERROR("add template specialization for this size\n");
GGML_ASSERT(false && "add template specialization for this size");
GGML_ABORT("add template specialization for this size");
}
}
} else {
@@ -2658,7 +2658,7 @@ static enum ggml_status ggml_metal_graph_compute(
{
GGML_METAL_LOG_ERROR("unsupported size: %lld\n", ne00);
GGML_METAL_LOG_ERROR("add template specialization for this size\n");
GGML_ASSERT(false && "add template specialization for this size");
GGML_ABORT("add template specialization for this size");
}
}
}
@@ -2779,7 +2779,7 @@ static enum ggml_status ggml_metal_graph_compute(
case GGML_TYPE_Q5_0: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_F32_Q5_0].pipeline; break;
case GGML_TYPE_Q5_1: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_F32_Q5_1].pipeline; break;
case GGML_TYPE_IQ4_NL: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_F32_IQ4_NL].pipeline; break;
default: GGML_ASSERT(false && "not implemented");
default: GGML_ABORT("not implemented");
};
} break;
case GGML_TYPE_F16:
@@ -2787,10 +2787,10 @@ static enum ggml_status ggml_metal_graph_compute(
switch (dstt) {
case GGML_TYPE_F32: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_F16_F32].pipeline; break;
case GGML_TYPE_F16: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_F16_F16].pipeline; break;
default: GGML_ASSERT(false && "not implemented");
default: GGML_ABORT("not implemented");
};
} break;
default: GGML_ASSERT(false && "not implemented");
default: GGML_ABORT("not implemented");
}
[encoder setComputePipelineState:pipeline];
@@ -2818,7 +2818,7 @@ static enum ggml_status ggml_metal_graph_compute(
default:
{
GGML_METAL_LOG_ERROR("%s: error: node %3d, op = %8s not implemented\n", __func__, i, ggml_op_name(dst->op));
GGML_ASSERT(false);
GGML_ABORT("fatal error");
}
}