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

metal : handle buffers larger than device's maxBufferLength (#1826)

Browse source 
* metal : handle buffers larger than device's maxBufferLength

* metal : print more verbose device info + handle errors

* metal : fix prints for overlapping views

* metal : minimize view overlap to try to utilize device memory better
This commit is contained in:
Georgi Gerganov 2023-06-18 09:09:47 +03:00 • committed by GitHub
parent 57cd69460f
commit ce2c7d72e2
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
6 changed files with 125 additions and 35 deletions
Download patch file Download diff file Expand all files Collapse all files

26
llama.cpp
View file

@ -2696,16 +2696,21 @@ struct llama_context * llama_init_from_file(
// this allocates all Metal resources and memory buffers
ctx->ctx_metal = ggml_metal_init();
void *data_ptr = NULL;
void * data_ptr = NULL;
size_t data_size = 0;
if (params.use_mmap) {
data_ptr = ctx->model.mapping->addr;
data_size= ctx->model.mapping->size;
data_ptr = ctx->model.mapping->addr;
data_size = ctx->model.mapping->size;
} else {
data_ptr = ggml_get_mem_buffer(ctx->model.ctx);
data_size= ggml_get_mem_size(ctx->model.ctx);
data_ptr = ggml_get_mem_buffer(ctx->model.ctx);
data_size = ggml_get_mem_size (ctx->model.ctx);
}
const size_t max_size = ggml_get_max_tensor_size(ctx->model.ctx);
printf("%s: max tensor size = %8.2f MB\n", __func__, max_size/1024.0/1024.0);
#define LLAMA_METAL_CHECK_BUF(result) \
if (!(result)) { \
fprintf(stderr, "%s: failed to add buffer\n", __func__); \
@ -2713,12 +2718,13 @@ struct llama_context * llama_init_from_file(
return NULL; \
}
LLAMA_METAL_CHECK_BUF(ggml_metal_add_buffer(ctx->ctx_metal, "data", data_ptr, data_size));
LLAMA_METAL_CHECK_BUF(ggml_metal_add_buffer(ctx->ctx_metal, "eval", ctx->buf_compute.addr, ctx->buf_compute.size));
LLAMA_METAL_CHECK_BUF(ggml_metal_add_buffer(ctx->ctx_metal, "data", data_ptr, data_size, max_size));
LLAMA_METAL_CHECK_BUF(ggml_metal_add_buffer(ctx->ctx_metal, "kv", ctx->model.kv_self.buf.addr, ctx->model.kv_self.buf.size));
LLAMA_METAL_CHECK_BUF(ggml_metal_add_buffer(ctx->ctx_metal, "scr0", ctx->buf_scratch[0].addr, ctx->buf_scratch[0].size));
LLAMA_METAL_CHECK_BUF(ggml_metal_add_buffer(ctx->ctx_metal, "scr1", ctx->buf_scratch[1].addr, ctx->buf_scratch[1].size));
LLAMA_METAL_CHECK_BUF(ggml_metal_add_buffer(ctx->ctx_metal, "eval", ctx->buf_compute.addr, ctx->buf_compute.size, 0));
LLAMA_METAL_CHECK_BUF(ggml_metal_add_buffer(ctx->ctx_metal, "kv", ctx->model.kv_self.buf.addr, ctx->model.kv_self.buf.size, 0));
LLAMA_METAL_CHECK_BUF(ggml_metal_add_buffer(ctx->ctx_metal, "scr0", ctx->buf_scratch[0].addr, ctx->buf_scratch[0].size, 0));
LLAMA_METAL_CHECK_BUF(ggml_metal_add_buffer(ctx->ctx_metal, "scr1", ctx->buf_scratch[1].addr, ctx->buf_scratch[1].size, 0));
#undef LLAMA_METAL_CHECK_BUF
}
#endif