vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

llama : sync with recent PRs on master

Browse source
This commit is contained in:
Georgi Gerganov 2023-08-17 19:03:15 +03:00
parent 8ace03ad3d
commit 11bf4366c2
No known key found for this signature in database
GPG key ID: 449E073F9DC10735
7 changed files with 616 additions and 672 deletions
Download patch file Download diff file Expand all files Collapse all files

195
ggml-metal.m
View file

@ -5,7 +5,6 @@
#import <Foundation/Foundation.h>
#import <Metal/Metal.h>
#import <MetalPerformanceShaders/MetalPerformanceShaders.h>
#undef MIN
#undef MAX
@ -79,6 +78,14 @@ struct ggml_metal_context {
GGML_METAL_DECL_KERNEL(mul_mat_q4_K_f32);
GGML_METAL_DECL_KERNEL(mul_mat_q5_K_f32);
GGML_METAL_DECL_KERNEL(mul_mat_q6_K_f32);
GGML_METAL_DECL_KERNEL(mul_mm_f16_f32);
GGML_METAL_DECL_KERNEL(mul_mm_q4_0_f32);
GGML_METAL_DECL_KERNEL(mul_mm_q4_1_f32);
GGML_METAL_DECL_KERNEL(mul_mm_q2_K_f32);
GGML_METAL_DECL_KERNEL(mul_mm_q3_K_f32);
GGML_METAL_DECL_KERNEL(mul_mm_q4_K_f32);
GGML_METAL_DECL_KERNEL(mul_mm_q5_K_f32);
GGML_METAL_DECL_KERNEL(mul_mm_q6_K_f32);
GGML_METAL_DECL_KERNEL(rope);
GGML_METAL_DECL_KERNEL(alibi_f32);
GGML_METAL_DECL_KERNEL(cpy_f32_f16);
@ -110,13 +117,6 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) {
ctx->n_buffers = 0;
ctx->concur_list_len = 0;
// determine if we can use MPS
if (MPSSupportsMTLDevice(ctx->device)) {
fprintf(stderr, "%s: using MPS\n", __func__);
} else {
fprintf(stderr, "%s: not using MPS\n", __func__);
GGML_ASSERT(false && "MPS not supported");
}
#if 0
// compile from source string and show compile log
@ -163,10 +163,15 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) {
// load kernels
{
NSError * error = nil;
#define GGML_METAL_ADD_KERNEL(name) \
ctx->function_##name = [ctx->library newFunctionWithName:@"kernel_"#name]; \
ctx->pipeline_##name = [ctx->device newComputePipelineStateWithFunction:ctx->function_##name error:nil]; \
fprintf(stderr, "%s: loaded %-32s %16p\n", __func__, "kernel_"#name, (void *) ctx->pipeline_##name);
ctx->pipeline_##name = [ctx->device newComputePipelineStateWithFunction:ctx->function_##name error:&error]; \
fprintf(stderr, "%s: loaded %-32s %16p\n", __func__, "kernel_"#name, (void *) ctx->pipeline_##name); \
if (error) { \
fprintf(stderr, "%s: load pipeline error: %s\n", __func__, [[error description] UTF8String]); \
return NULL; \
}
GGML_METAL_ADD_KERNEL(add);
GGML_METAL_ADD_KERNEL(add_row);
@ -196,6 +201,14 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) {
GGML_METAL_ADD_KERNEL(mul_mat_q4_K_f32);
GGML_METAL_ADD_KERNEL(mul_mat_q5_K_f32);
GGML_METAL_ADD_KERNEL(mul_mat_q6_K_f32);
GGML_METAL_ADD_KERNEL(mul_mm_f16_f32);
GGML_METAL_ADD_KERNEL(mul_mm_q4_0_f32);
GGML_METAL_ADD_KERNEL(mul_mm_q4_1_f32);
GGML_METAL_ADD_KERNEL(mul_mm_q2_K_f32);
GGML_METAL_ADD_KERNEL(mul_mm_q3_K_f32);
GGML_METAL_ADD_KERNEL(mul_mm_q4_K_f32);
GGML_METAL_ADD_KERNEL(mul_mm_q5_K_f32);
GGML_METAL_ADD_KERNEL(mul_mm_q6_K_f32);
GGML_METAL_ADD_KERNEL(rope);
GGML_METAL_ADD_KERNEL(alibi_f32);
GGML_METAL_ADD_KERNEL(cpy_f32_f16);
@ -243,11 +256,12 @@ void ggml_metal_set_n_cb(struct ggml_metal_context * ctx, int n_cb) {
ctx->n_cb = n_cb;
}
bool ggml_metal_if_optimized(struct ggml_metal_context * ctx) {
if (ctx->concur_list_len) {
return true;
}
return false;
int ggml_metal_if_optimized(struct ggml_metal_context * ctx) {
return ctx->concur_list_len;
}
int * ggml_metal_get_concur_list(struct ggml_metal_context * ctx) {
return ctx->concur_list;
}
// finds the Metal buffer that contains the tensor data on the GPU device
@ -390,7 +404,7 @@ void ggml_metal_get_tensor(
void ggml_metal_graph_find_concurrency(
struct ggml_metal_context * ctx,
struct ggml_cgraph * gf) {
struct ggml_cgraph * gf, bool check_mem) {
int search_depth = gf->n_nodes; //we only find concurrency in this range to avoid wasting too much time
int nodes_unused[GGML_MAX_CONCUR];
@ -437,7 +451,7 @@ void ggml_metal_graph_find_concurrency(
}
}
}
if (exe_flag) {
if (exe_flag && check_mem) {
// check if nodes[i]'s data will be overwritten by a node before nodes[i].
// if node[5] and node[3] write to the same memory region, then we can't issue node[5] before node[3]
int64_t data_start = (int64_t) gf->nodes[i]->data;
@ -521,7 +535,7 @@ void ggml_metal_graph_compute(
id<MTLCommandBuffer> command_buffer = command_buffers[cb_idx];
id<MTLComputeCommandEncoder> encoder = nil;
id<MTLComputeCommandEncoder> encoder = [command_buffer computeCommandEncoderWithDescriptor: edesc];
const int node_start = (cb_idx + 0) * n_nodes_per_cb;
const int node_end = (cb_idx == n_cb - 1) ? n_nodes : (cb_idx + 1) * n_nodes_per_cb;
@ -530,10 +544,6 @@ void ggml_metal_graph_compute(
const int i = has_concur ? ctx->concur_list[ind] : ind;
if (i == -1) {
if (encoder == nil) {
encoder = [command_buffer computeCommandEncoderWithDescriptor: edesc];
continue;
}
[encoder memoryBarrierWithScope:MTLBarrierScopeBuffers];
continue;
}
@ -607,10 +617,6 @@ void ggml_metal_graph_compute(
} break;
case GGML_OP_ADD:
{
if (encoder == nil) {
encoder = [command_buffer computeCommandEncoderWithDescriptor: edesc];
}
if (ggml_nelements(src1) == ne10) {
// src1 is a row
[encoder setComputePipelineState:ctx->pipeline_add_row];
@ -628,10 +634,6 @@ void ggml_metal_graph_compute(
} break;
case GGML_OP_MUL:
{
if (encoder == nil) {
encoder = [command_buffer computeCommandEncoderWithDescriptor: edesc];
}
if (ggml_nelements(src1) == ne10) {
// src1 is a row
[encoder setComputePipelineState:ctx->pipeline_mul_row];
@ -649,10 +651,6 @@ void ggml_metal_graph_compute(
} break;
case GGML_OP_SCALE:
{
if (encoder == nil) {
encoder = [command_buffer computeCommandEncoderWithDescriptor: edesc];
}
const float scale = *(const float *) src1->data;
[encoder setComputePipelineState:ctx->pipeline_scale];
@ -668,10 +666,6 @@ void ggml_metal_graph_compute(
switch (ggml_get_unary_op(gf->nodes[i])) {
case GGML_UNARY_OP_SILU:
{
if (encoder == nil) {
encoder = [command_buffer computeCommandEncoderWithDescriptor: edesc];
}
[encoder setComputePipelineState:ctx->pipeline_silu];
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
[encoder setBuffer:id_dst offset:offs_dst atIndex:1];
@ -682,10 +676,6 @@ void ggml_metal_graph_compute(
} break;
case GGML_UNARY_OP_RELU:
{
if (encoder == nil) {
encoder = [command_buffer computeCommandEncoderWithDescriptor: edesc];
}
[encoder setComputePipelineState:ctx->pipeline_relu];
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
[encoder setBuffer:id_dst offset:offs_dst atIndex:1];
@ -696,10 +686,6 @@ void ggml_metal_graph_compute(
} break;
case GGML_UNARY_OP_GELU:
{
if (encoder == nil) {
encoder = [command_buffer computeCommandEncoderWithDescriptor: edesc];
}
[encoder setComputePipelineState:ctx->pipeline_gelu];
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
[encoder setBuffer:id_dst offset:offs_dst atIndex:1];
@ -716,10 +702,6 @@ void ggml_metal_graph_compute(
} break;
case GGML_OP_SOFT_MAX:
{
if (encoder == nil) {
encoder = [command_buffer computeCommandEncoderWithDescriptor: edesc];
}
const int nth = 32;
[encoder setComputePipelineState:ctx->pipeline_soft_max];
@ -734,10 +716,6 @@ void ggml_metal_graph_compute(
} break;
case GGML_OP_DIAG_MASK_INF:
{
if (encoder == nil) {
encoder = [command_buffer computeCommandEncoderWithDescriptor: edesc];
}
const int n_past = ((int32_t *)(dst->op_params))[0];
[encoder setComputePipelineState:ctx->pipeline_diag_mask_inf];
@ -755,53 +733,43 @@ void ggml_metal_graph_compute(
GGML_ASSERT(ne00 == ne10);
// GGML_ASSERT(ne02 == ne12); // Should be checked on individual data types until broadcast is implemented everywhere
uint gqa = ne12/ne02;
GGML_ASSERT(ne03 == ne13);
// for now the matrix-matrix multiplication kernel only works on A14+/M1+ SoCs
// AMD GPU and older A-chips will reuse matrix-vector multiplication kernel
if (ggml_is_contiguous(src0) &&
ggml_is_contiguous(src1) &&
(src0t == GGML_TYPE_F32 || src0t == GGML_TYPE_F16) && ne11 > 1) {
if (encoder != nil) {
[encoder endEncoding];
encoder = nil;
src1t == GGML_TYPE_F32 &&
[ctx->device supportsFamily:MTLGPUFamilyApple7] &&
ne00%32 == 0 &&
ne11 > 1) {
switch (src0->type) {
case GGML_TYPE_F16: [encoder setComputePipelineState:ctx->pipeline_mul_mm_f16_f32]; break;
case GGML_TYPE_Q4_0: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q4_0_f32]; break;
case GGML_TYPE_Q4_1: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q4_1_f32]; break;
case GGML_TYPE_Q2_K: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q2_K_f32]; break;
case GGML_TYPE_Q3_K: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q3_K_f32]; break;
case GGML_TYPE_Q4_K: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q4_K_f32]; break;
case GGML_TYPE_Q5_K: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q5_K_f32]; break;
case GGML_TYPE_Q6_K: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q6_K_f32]; break;
default: GGML_ASSERT(false && "MUL MAT-MAT not implemented");
}
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
[encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
[encoder setBuffer:id_dst offset:offs_dst atIndex:2];
[encoder setBytes:&ne00 length:sizeof(ne00) atIndex:3];
[encoder setBytes:&ne02 length:sizeof(ne02) atIndex:4];
[encoder setBytes:&nb01 length:sizeof(nb01) atIndex:5];
[encoder setBytes:&nb02 length:sizeof(nb02) atIndex:6];
[encoder setBytes:&ne12 length:sizeof(ne12) atIndex:7];
[encoder setBytes:&ne0 length:sizeof(ne0) atIndex:8];
[encoder setBytes:&ne1 length:sizeof(ne1) atIndex:9];
[encoder setBytes:&gqa length:sizeof(gqa) atIndex:10];
[encoder setThreadgroupMemoryLength:8192 atIndex:0];
[encoder dispatchThreadgroups:MTLSizeMake( (ne11+31)/32, (ne01+63) / 64, ne12) threadsPerThreadgroup:MTLSizeMake(128, 1, 1)];
}
MPSDataType src0dt = src0t == GGML_TYPE_F32 ? MPSDataTypeFloat32 : MPSDataTypeFloat16;
MPSDataType src1dt = src1t == GGML_TYPE_F32 ? MPSDataTypeFloat32 : MPSDataTypeFloat16;
// for F32 x F32 we use MPS
MPSMatrixDescriptor * desc0 = [MPSMatrixDescriptor
matrixDescriptorWithRows:ne01 columns:ne00 rowBytes:src0->nb[1] dataType:src0dt];
MPSMatrixDescriptor * desc1 = [MPSMatrixDescriptor
matrixDescriptorWithRows:ne11 columns:ne10 rowBytes:src1->nb[1] dataType:src1dt];
MPSMatrixDescriptor * desc = [MPSMatrixDescriptor
matrixDescriptorWithRows:ne1 columns:ne0 rowBytes:dst->nb[1] dataType:MPSDataTypeFloat32];
MPSMatrixMultiplication * mul = [[MPSMatrixMultiplication alloc]
initWithDevice:ctx->device transposeLeft:false transposeRight:true
resultRows:ne11 resultColumns:ne01 interiorColumns:ne00 alpha:1.0 beta:0.0];
// we need to do ne12 multiplications
// TODO: is there a way to do this in parallel - currently very slow ..
// TODO: might be possible to offload part of the computation to ANE using Accelerate's CBLAS
for (int64_t i02 = 0; i02 < ne12; ++i02) {
size_t offs_src0_cur = offs_src0 + i02/(ne12/ne02)*nb02; // gqa not used for now
size_t offs_src1_cur = offs_src1 + i02*nb12;
size_t offs_dst_cur = offs_dst + i02*nb2;
MPSMatrix * mat_src0 = [[MPSMatrix alloc] initWithBuffer:id_src0 offset:offs_src0_cur descriptor:desc0];
MPSMatrix * mat_src1 = [[MPSMatrix alloc] initWithBuffer:id_src1 offset:offs_src1_cur descriptor:desc1];
MPSMatrix * mat_dst = [[MPSMatrix alloc] initWithBuffer:id_dst offset:offs_dst_cur descriptor:desc ];
[mul encodeToCommandBuffer:command_buffer leftMatrix:mat_src1 rightMatrix:mat_src0 resultMatrix:mat_dst];
}
} else {
if (encoder == nil) {
encoder = [command_buffer computeCommandEncoderWithDescriptor: edesc];
}
else {
int nth0 = 32;
int nth1 = 1;
@ -900,23 +868,24 @@ void ggml_metal_graph_compute(
[encoder setBytes:&nb12 length:sizeof(nb12) atIndex:14];
[encoder setBytes:&ne0 length:sizeof(ne0) atIndex:15];
[encoder setBytes:&ne1 length:sizeof(ne1) atIndex:16];
[encoder setBytes:&gqa length:sizeof(gqa) atIndex:17];
if (src0t == GGML_TYPE_Q4_0 || src0t == GGML_TYPE_Q4_1 ||
src0t == GGML_TYPE_Q2_K || src0t == GGML_TYPE_Q4_K) {
[encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7) / 8, ne11, 1) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
[encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7) / 8, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
}
else if (src0t == GGML_TYPE_Q3_K) {
#ifdef GGML_QKK_64
[encoder dispatchThreadgroups:MTLSizeMake((ne01+1)/2, ne11, 1) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
[encoder dispatchThreadgroups:MTLSizeMake((ne01+1)/2, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
#else
[encoder dispatchThreadgroups:MTLSizeMake((ne01+3)/4, ne11, 1) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
[encoder dispatchThreadgroups:MTLSizeMake((ne01+3)/4, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
#endif
}
else if (src0t == GGML_TYPE_Q5_K) {
[encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3) / 4, ne11, 1) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
[encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3) / 4, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
}
else if (src0t == GGML_TYPE_Q6_K) {
[encoder dispatchThreadgroups:MTLSizeMake((ne01+1)/2, ne11, 1) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
[encoder dispatchThreadgroups:MTLSizeMake((ne01+1)/2, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
} else {
[encoder setThreadgroupMemoryLength:nth0*sizeof(float) atIndex:0];
[encoder dispatchThreadgroups:MTLSizeMake(ne01, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
@ -925,10 +894,6 @@ void ggml_metal_graph_compute(
} break;
case GGML_OP_GET_ROWS:
{
if (encoder == nil) {
encoder = [command_buffer computeCommandEncoderWithDescriptor: edesc];
}
switch (src0->type) {
case GGML_TYPE_F16: [encoder setComputePipelineState:ctx->pipeline_get_rows_f16]; break;
case GGML_TYPE_Q4_0: [encoder setComputePipelineState:ctx->pipeline_get_rows_q4_0]; break;
@ -954,10 +919,6 @@ void ggml_metal_graph_compute(
} break;
case GGML_OP_RMS_NORM:
{
if (encoder == nil) {
encoder = [command_buffer computeCommandEncoderWithDescriptor: edesc];
}
float eps;
memcpy(&eps, dst->op_params, sizeof(float));
@ -977,10 +938,6 @@ void ggml_metal_graph_compute(
} break;
case GGML_OP_NORM:
{
if (encoder == nil) {
encoder = [command_buffer computeCommandEncoderWithDescriptor: edesc];
}
const float eps = 1e-5f;
const int nth = 256;
@ -999,10 +956,6 @@ void ggml_metal_graph_compute(
} break;
case GGML_OP_ALIBI:
{
if (encoder == nil) {
encoder = [command_buffer computeCommandEncoderWithDescriptor: edesc];
}
GGML_ASSERT((src0t == GGML_TYPE_F32));
const int n_past = ((int32_t *) dst->op_params)[0]; UNUSED(n_past);
@ -1042,10 +995,6 @@ void ggml_metal_graph_compute(
} break;
case GGML_OP_ROPE:
{
if (encoder == nil) {
encoder = [command_buffer computeCommandEncoderWithDescriptor: edesc];
}
const int n_past = ((int32_t *) dst->op_params)[0];
const int n_dims = ((int32_t *) dst->op_params)[1];
const int mode = ((int32_t *) dst->op_params)[2];
@ -1086,10 +1035,6 @@ void ggml_metal_graph_compute(
case GGML_OP_CPY:
case GGML_OP_CONT:
{
if (encoder == nil) {
encoder = [command_buffer computeCommandEncoderWithDescriptor: edesc];
}
const int nth = 32;
switch (src0t) {