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

convert-hf : support bfloat16 conversion (#7158)

Browse source 
* convert-hf : support bfloat16 conversion

* gguf-py : flake8 fixes

* convert-hf : add missing space after comma

* convert-hf : get bit-exact same output as ./quantize

The quantization version was missing.

* convert-hf : don't round bf16 NANs

* convert-hf : save some memory with np.int16 intermediate bf16 weights

* convert-hf : more closely match llama.cpp with which weights to keep in f32

* convert-hf : add --outtype auto-f16

A reason for this to exist is for model quantizers who want an initial
GGUF with the most fidelity to the original model while still using
a 16-bit float type instead of 32-bit floats.

* convert-hf : remove a semicolon because flake8 doesn't like it

It's a reflex from when programming in C/C++, I guess.

* convert-hf : support outtype templating in outfile name

* convert-hf : rename --outtype auto-f16 to --outtype auto
This commit is contained in:
compilade 2024-05-11 11:06:26 -04:00 • committed by GitHub
parent fae9d234b6
commit 5a419926b0
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
5 changed files with 404 additions and 182 deletions
Download patch file Download diff file Expand all files Collapse all files

44
gguf-py/gguf/constants.py
View file

@ -10,6 +10,7 @@ from typing import Any
GGUF_MAGIC = 0x46554747 # "GGUF"
GGUF_VERSION = 3
GGUF_DEFAULT_ALIGNMENT = 32
GGML_QUANT_VERSION = 2 # GGML_QNT_VERSION from ggml.h
#
# metadata keys
@ -838,6 +839,49 @@ class GGMLQuantizationType(IntEnum):
BF16 = 30
# TODO: add GGMLFileType from ggml_ftype in ggml.h
# from llama_ftype in llama.h
# ALL VALUES SHOULD BE THE SAME HERE AS THEY ARE OVER THERE.
class LlamaFileType(IntEnum):
ALL_F32 = 0
MOSTLY_F16 = 1 # except 1d tensors
MOSTLY_Q4_0 = 2 # except 1d tensors
MOSTLY_Q4_1 = 3 # except 1d tensors
MOSTLY_Q4_1_SOME_F16 = 4 # tok_embeddings.weight and output.weight are F16
# MOSTLY_Q4_2 = 5 # support has been removed
# MOSTLY_Q4_3 = 6 # support has been removed
MOSTLY_Q8_0 = 7 # except 1d tensors
MOSTLY_Q5_0 = 8 # except 1d tensors
MOSTLY_Q5_1 = 9 # except 1d tensors
MOSTLY_Q2_K = 10 # except 1d tensors
MOSTLY_Q3_K_S = 11 # except 1d tensors
MOSTLY_Q3_K_M = 12 # except 1d tensors
MOSTLY_Q3_K_L = 13 # except 1d tensors
MOSTLY_Q4_K_S = 14 # except 1d tensors
MOSTLY_Q4_K_M = 15 # except 1d tensors
MOSTLY_Q5_K_S = 16 # except 1d tensors
MOSTLY_Q5_K_M = 17 # except 1d tensors
MOSTLY_Q6_K = 18 # except 1d tensors
MOSTLY_IQ2_XXS = 19 # except 1d tensors
MOSTLY_IQ2_XS = 20 # except 1d tensors
MOSTLY_Q2_K_S = 21 # except 1d tensors
MOSTLY_IQ3_XS = 22 # except 1d tensors
MOSTLY_IQ3_XXS = 23 # except 1d tensors
MOSTLY_IQ1_S = 24 # except 1d tensors
MOSTLY_IQ4_NL = 25 # except 1d tensors
MOSTLY_IQ3_S = 26 # except 1d tensors
MOSTLY_IQ3_M = 27 # except 1d tensors
MOSTLY_IQ2_S = 28 # except 1d tensors
MOSTLY_IQ2_M = 29 # except 1d tensors
MOSTLY_IQ4_XS = 30 # except 1d tensors
MOSTLY_IQ1_M = 31 # except 1d tensors
MOSTLY_BF16 = 32 # except 1d tensors
GUESSED = 1024 # not specified in the model file
class GGUFEndian(IntEnum):
LITTLE = 0
BIG = 1