feat: Update llama.cpp

abetlen · Sep 29, 2024 · 57e70bb · 57e70bb
1 parent 01c7607
commit 57e70bb
Show file tree

Hide file tree

Showing 2 changed files with 13 additions and 5 deletions.
diff --git a/llama_cpp/llama_cpp.py b/llama_cpp/llama_cpp.py
@@ -220,6 +220,7 @@
 #     LLAMA_VOCAB_PRE_TYPE_BLOOM          = 23,
 #     LLAMA_VOCAB_PRE_TYPE_GPT3_FINNISH   = 24,
 #     LLAMA_VOCAB_PRE_TYPE_EXAONE         = 25,
+#     LLAMA_VOCAB_PRE_TYPE_CHAMELEON      = 26,
 # };
 LLAMA_VOCAB_PRE_TYPE_DEFAULT = 0
 LLAMA_VOCAB_PRE_TYPE_LLAMA3 = 1
@@ -247,6 +248,7 @@
 LLAMA_VOCAB_PRE_TYPE_BLOOM = 23
 LLAMA_VOCAB_PRE_TYPE_GPT3_FINNISH = 24
 LLAMA_VOCAB_PRE_TYPE_EXAONE = 25
+LLAMA_VOCAB_PRE_TYPE_CHAMELEON = 26
 
 
 # // note: these values should be synchronized with ggml_rope
@@ -404,12 +406,14 @@
 #     LLAMA_POOLING_TYPE_MEAN = 1,
 #     LLAMA_POOLING_TYPE_CLS  = 2,
 #     LLAMA_POOLING_TYPE_LAST = 3,
+#     LLAMA_POOLING_TYPE_RANK = 4, // used by reranking models to attach the classification head to the graph
 # };
 LLAMA_POOLING_TYPE_UNSPECIFIED = -1
 LLAMA_POOLING_TYPE_NONE = 0
 LLAMA_POOLING_TYPE_MEAN = 1
 LLAMA_POOLING_TYPE_CLS = 2
 LLAMA_POOLING_TYPE_LAST = 3
+LLAMA_POOLING_TYPE_RANK = 4
 
 # enum llama_attention_type {
 #     LLAMA_ATTENTION_TYPE_UNSPECIFIED = -1,
@@ -420,10 +424,11 @@
 LLAMA_ATTENTION_TYPE_CAUSAL = 0
 LLAMA_ATTENTION_TYPE_NON_CAUSAL = 1
 
+
 # enum llama_split_mode {
-#     LLAMA_SPLIT_MODE_NONE    = 0, // single GPU
-#     LLAMA_SPLIT_MODE_LAYER   = 1, // split layers and KV across GPUs
-#     LLAMA_SPLIT_MODE_ROW     = 2, // split rows across GPUs
+#     LLAMA_SPLIT_MODE_NONE  = 0, // single GPU
+#     LLAMA_SPLIT_MODE_LAYER = 1, // split layers and KV across GPUs
+#     LLAMA_SPLIT_MODE_ROW   = 2, // split rows across GPUs
 # };
 LLAMA_SPLIT_MODE_NONE = 0
 LLAMA_SPLIT_MODE_LAYER = 1
@@ -2520,7 +2525,8 @@ def llama_get_embeddings_ith(
 
 # // Get the embeddings for a sequence id
 # // Returns NULL if pooling_type is LLAMA_POOLING_TYPE_NONE
-# // shape: [n_embd] (1-dimensional)
+# // when pooling_type == LLAMA_POOLING_TYPE_RANK, returns float[1] with the rank of the sequence
+# // otherwise: float[n_embd] (1-dimensional)
 # LLAMA_API float * llama_get_embeddings_seq(struct llama_context * ctx, llama_seq_id seq_id);
 @ctypes_function(
     "llama_get_embeddings_seq",
@@ -2672,6 +2678,8 @@ def llama_token_eot(model: llama_model_p, /) -> int:
 # //
 # // Tokenization
 # //
+# // The API is thread-safe.
+# //
 
 
 # /// @details Convert the provided text into tokens.

diff --git a/vendor/llama.cpp b/vendor/llama.cpp
+8 −3		CMakeLists.txt
+5 −0		CONTRIBUTING.md
+4 −2		README.md
+76 −9		ci/run.sh
+15 −3		common/arg.cpp
+7 −0		common/common.cpp
+1 −0		common/common.h
+69 −3		convert_hf_to_gguf.py
+2 −0		convert_hf_to_gguf_update.py
+7 −7		examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp
+6 −1		examples/embedding/embedding.cpp
+38 −1		examples/server/README.md
+199 −16		examples/server/server.cpp
+1 −1		examples/server/tests/features/embeddings.feature
+42 −0		examples/server/tests/features/rerank.feature
+53 −1		examples/server/tests/features/steps/steps.py
+24 −1		examples/server/utils.hpp
+15 −10		ggml/include/ggml.h
+2 −11		ggml/src/ggml-aarch64.c
+0 −1		ggml/src/ggml-cuda/im2col.cu
+2 −2		ggml/src/ggml-quants.c
+0 −4		ggml/src/ggml-quants.h
+57 −12		ggml/src/ggml-vulkan.cpp
+90 −15		ggml/src/ggml.c
+4 −6		ggml/src/vulkan-shaders/argsort.comp
+26 −0		gguf-py/gguf/constants.py
+3 −0		gguf-py/gguf/gguf_writer.py
+11 −2		gguf-py/gguf/tensor_mapping.py
+9 −4		include/llama.h
+112 −0		models/ggml-vocab-chameleon.gguf.inp
+46 −0		models/ggml-vocab-chameleon.gguf.out
+1 −1		scripts/sync-ggml.last
+189 −106		src/llama-vocab.cpp
+9 −0		src/llama-vocab.h
+351 −10		src/llama.cpp
+216 −199		tests/test-backend-ops.cpp
+55 −35		tests/test-tokenizer-0.cpp