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

ggml-alloc: avoid return silently

Browse source 
In certain cases, the allocate_node() function may silently return
without performing any memory allocation.
This commit is contained in:
lshzh-ww 2023-08-24 01:34:57 -04:00
parent ee8b2aa75d
commit 0c268a83e8
2 changed files with 2 additions and 7 deletions
Download patch file Download diff file Expand all files Collapse all files

2
ggml-alloc.c
View file

@ -441,11 +441,11 @@ static void allocate_node(struct ggml_allocr * alloc, struct ggml_tensor * node)
else { else {
AT_PRINTF("reusing parent %s for %s\n", parent->name, node->name); AT_PRINTF("reusing parent %s for %s\n", parent->name, node->name);
node->data = parent->data; node->data = parent->data;
}
return; return;
} }
} }
} }
}
ggml_allocr_alloc(alloc, node); ggml_allocr_alloc(alloc, node);
} }
} }

5
llama.cpp
View file

@ -2704,11 +2704,6 @@ static struct ggml_cgraph * llm_build_falcon(
struct ggml_tensor * inpFF = attn_norm; struct ggml_tensor * inpFF = attn_norm;
cur = ggml_mul_mat(ctx0, model.layers[il].w3, inpFF); cur = ggml_mul_mat(ctx0, model.layers[il].w3, inpFF);
// TODO: this is temporary needed to introduce artificial dependency between FF and ATTN
// adding this, because there seems to be a bug in the Metal concurrency optimization
// without this line, the results are non-deterministic and wrong
cur->src[2] = attn_out;
offload_func(cur); offload_func(cur);
cur = ggml_gelu(ctx0, cur); cur = ggml_gelu(ctx0, cur);