matt/ollama37
1
0
Fork 0
mirror of https://github.com/dogkeeper886/ollama37.git synced 2025-12-14 17:57:06 +00:00
Code Issues Packages Projects Releases Wiki Activity

llama: update vendor code to commit ba1cb19c (#8101)

Browse Source
This commit is contained in:
Jeffrey Morgan
2024-12-14 14:55:51 -08:00
committed by GitHub
parent 60f75560a2
commit 7a81daf026
273 changed files with 3194 additions and 1900 deletions
Download Patch File Download Diff File Expand all files Collapse all files

79
llama/patches/0009-add-unpad-operator.patch
View File

@@ -5,30 +5,30 @@ Subject: [PATCH] add unpad operator
---
ggml/include/ggml.h | 10 +++++
ggml/src/ggml-cpu/ggml-cpu.c | 57 ++++++++++++++++++++++++++++
ggml/src/ggml-cpu/ggml-cpu.c | 58 ++++++++++++++++++++++++++++
ggml/src/ggml-cuda/ggml-cuda.cu | 4 ++
ggml/src/ggml-cuda/pad.cu | 46 ++++++++++++++++++++++
ggml/src/ggml-cuda/pad.cuh | 1 +
ggml/src/ggml-metal/ggml-metal.m | 33 ++++++++++++++++
ggml/src/ggml-metal/ggml-metal.metal | 45 ++++++++++++++++++++++
ggml/src/ggml-metal/ggml-metal.metal | 45 +++++++++++++++++++++
ggml/src/ggml.c | 25 +++++++++++-
8 files changed, 219 insertions(+), 2 deletions(-)
8 files changed, 220 insertions(+), 2 deletions(-)
diff --git a/ggml/include/ggml.h b/ggml/include/ggml.h
index 65cb92c4..acbcccc6 100644
index b0c1ac9c..091e6e6b 100644
--- a/ggml/include/ggml.h
+++ b/ggml/include/ggml.h
@@ -499,6 +499,7 @@ extern "C" {
GGML_OP_POOL_2D_BACK,
GGML_OP_UPSCALE, // nearest interpolate
GGML_OP_PAD,
GGML_OP_PAD_REFLECT_1D,
+ GGML_OP_UNPAD,
GGML_OP_ARANGE,
GGML_OP_TIMESTEP_EMBEDDING,
GGML_OP_ARGSORT,
@@ -1695,6 +1696,15 @@ extern "C" {
int p2,
int p3);
@@ -1718,6 +1719,15 @@ extern "C" {
int p0,
int p1);
+ // unpad each dimension: [x, ..., x, y, ..., y] -> [x, ..., x]
+ GGML_API struct ggml_tensor * ggml_unpad(
@@ -43,10 +43,10 @@ index 65cb92c4..acbcccc6 100644
// timesteps: [N,]
// return: [N, dim]
diff --git a/ggml/src/ggml-cpu/ggml-cpu.c b/ggml/src/ggml-cpu/ggml-cpu.c
index 23ae2e10..111ff3b0 100644
index 67e67a08..bebff207 100644
--- a/ggml/src/ggml-cpu/ggml-cpu.c
+++ b/ggml/src/ggml-cpu/ggml-cpu.c
@@ -10439,6 +10439,58 @@ static void ggml_compute_forward_pad(
@@ -10588,6 +10588,59 @@ static void ggml_compute_forward_pad_reflect_1d(
}
}
@@ -102,12 +102,13 @@ index 23ae2e10..111ff3b0 100644
+ }
+ }
+}
+
// ggml_compute_forward_arange
@@ -12535,6 +12587,10 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
static void ggml_compute_forward_arange_f32(
@@ -12690,6 +12743,10 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
{
ggml_compute_forward_pad(params, tensor);
ggml_compute_forward_pad_reflect_1d(params, tensor);
} break;
+ case GGML_OP_UNPAD:
+ {
@@ -116,16 +117,16 @@ index 23ae2e10..111ff3b0 100644
case GGML_OP_ARANGE:
{
ggml_compute_forward_arange(params, tensor);
@@ -12877,6 +12933,7 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
} break;
@@ -13033,6 +13090,7 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
case GGML_OP_UPSCALE:
case GGML_OP_PAD:
case GGML_OP_PAD_REFLECT_1D:
+ case GGML_OP_UNPAD:
case GGML_OP_ARANGE:
case GGML_OP_TIMESTEP_EMBEDDING:
case GGML_OP_ARGSORT:
diff --git a/ggml/src/ggml-cuda/ggml-cuda.cu b/ggml/src/ggml-cuda/ggml-cuda.cu
index cbf4fddf..9ca6cb77 100644
index 8fd7c1a3..7c351b89 100644
--- a/ggml/src/ggml-cuda/ggml-cuda.cu
+++ b/ggml/src/ggml-cuda/ggml-cuda.cu
@@ -2085,6 +2085,9 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
@@ -210,34 +211,34 @@ index 8fd386b0..e2ededc3 100644
void ggml_cuda_op_pad(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+void ggml_cuda_op_unpad(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
diff --git a/ggml/src/ggml-metal/ggml-metal.m b/ggml/src/ggml-metal/ggml-metal.m
index 093ae900..cb9a1307 100644
index 28f590f9..787fc713 100644
--- a/ggml/src/ggml-metal/ggml-metal.m
+++ b/ggml/src/ggml-metal/ggml-metal.m
@@ -310,6 +310,7 @@ static void ggml_backend_metal_device_rel(struct ggml_backend_metal_device_conte
GGML_METAL_KERNEL_TYPE_CONV_TRANSPOSE_1D_F16_F32,
@@ -311,6 +311,7 @@ static void ggml_backend_metal_device_rel(struct ggml_backend_metal_device_conte
GGML_METAL_KERNEL_TYPE_UPSCALE_F32,
GGML_METAL_KERNEL_TYPE_PAD_F32,
GGML_METAL_KERNEL_TYPE_PAD_REFLECT_1D_F32,
+ GGML_METAL_KERNEL_TYPE_UNPAD_F32,
GGML_METAL_KERNEL_TYPE_ARANGE_F32,
GGML_METAL_KERNEL_TYPE_TIMESTEP_EMBEDDING_F32,
GGML_METAL_KERNEL_TYPE_ARGSORT_F32_I32_ASC,
@@ -877,6 +878,7 @@ @implementation GGMLMetalClass
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CONV_TRANSPOSE_1D_F16_F32, conv_transpose_1d_f16_f32, true);
@@ -910,6 +911,7 @@ @implementation GGMLMetalClass
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_UPSCALE_F32, upscale_f32, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_PAD_F32, pad_f32, true);
+ GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_UNPAD_F32, unpad_f32, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_PAD_REFLECT_1D_F32, pad_reflect_1d_f32, true);
+ GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_UNPAD_F32, unpad_f32, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_TIMESTEP_EMBEDDING_F32, timestep_embedding_f32, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ARANGE_F32, arange_f32, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ARGSORT_F32_I32_ASC, argsort_f32_i32_asc, true);
@@ -1099,6 +1101,7 @@ static bool ggml_metal_supports_op(const struct ggml_backend_metal_device_contex
case GGML_OP_POOL_2D:
@@ -1145,6 +1147,7 @@ static bool ggml_metal_supports_op(const struct ggml_backend_metal_device_contex
case GGML_OP_UPSCALE:
case GGML_OP_PAD:
case GGML_OP_PAD_REFLECT_1D:
+ case GGML_OP_UNPAD:
case GGML_OP_ARANGE:
case GGML_OP_TIMESTEP_EMBEDDING:
case GGML_OP_ARGSORT:
@@ -3258,6 +3261,36 @@ static void ggml_metal_encode_node(
@@ -3348,6 +3351,36 @@ static void ggml_metal_encode_node(
const int nth = MIN(1024, ne0);
@@ -275,10 +276,10 @@ index 093ae900..cb9a1307 100644
} break;
case GGML_OP_ARANGE:
diff --git a/ggml/src/ggml-metal/ggml-metal.metal b/ggml/src/ggml-metal/ggml-metal.metal
index 5caa0846..47038c31 100644
index 8ba43904..204c93e6 100644
--- a/ggml/src/ggml-metal/ggml-metal.metal
+++ b/ggml/src/ggml-metal/ggml-metal.metal
@@ -2897,6 +2897,51 @@ kernel void kernel_pad_f32(
@@ -2944,6 +2944,51 @@ kernel void kernel_pad_reflect_1d_f32(
}
}
@@ -331,44 +332,44 @@ index 5caa0846..47038c31 100644
device char * dst,
constant int64_t & ne0,
diff --git a/ggml/src/ggml.c b/ggml/src/ggml.c
index 1a9a7efa..ea2b259b 100644
index 51cc8566..0e74e554 100644
--- a/ggml/src/ggml.c
+++ b/ggml/src/ggml.c
@@ -950,6 +950,7 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = {
"POOL_2D_BACK",
@@ -954,6 +954,7 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = {
"UPSCALE",
"PAD",
"PAD_REFLECT_1D",
+ "UNPAD",
"ARANGE",
"TIMESTEP_EMBEDDING",
"ARGSORT",
@@ -983,7 +984,7 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = {
@@ -987,7 +988,7 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = {
"OPT_STEP_ADAMW",
};
-static_assert(GGML_OP_COUNT == 81, "GGML_OP_COUNT != 81");
+static_assert(GGML_OP_COUNT == 82, "GGML_OP_COUNT != 82");
-static_assert(GGML_OP_COUNT == 82, "GGML_OP_COUNT != 82");
+static_assert(GGML_OP_COUNT == 83, "GGML_OP_COUNT != 83");
static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
"none",
@@ -1045,6 +1046,7 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
"pool_2d_back(x)",
@@ -1050,6 +1051,7 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
"upscale(x)",
"pad(x)",
"pad_reflect_1d(x)",
+ "unpad(x)",
"arange(start, stop, step)",
"timestep_embedding(timesteps, dim, max_period)",
"argsort(x)",
@@ -1078,7 +1080,7 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
@@ -1083,7 +1085,7 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
"adamw(x)",
};
-static_assert(GGML_OP_COUNT == 81, "GGML_OP_COUNT != 81");
+static_assert(GGML_OP_COUNT == 82, "GGML_OP_COUNT != 82");
-static_assert(GGML_OP_COUNT == 82, "GGML_OP_COUNT != 82");
+static_assert(GGML_OP_COUNT == 83, "GGML_OP_COUNT != 83");
static_assert(GGML_OP_POOL_COUNT == 2, "GGML_OP_POOL_COUNT != 2");
@@ -4097,6 +4099,25 @@ struct ggml_tensor * ggml_pad(
@@ -4180,6 +4182,25 @@ struct ggml_tensor * ggml_pad_reflect_1d(
return result;
}