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

convert_hf : simplify internal quantization type selection

Browse source
This commit is contained in:
Francis Couture-Harpin 2024-08-02 16:14:49 -04:00
parent 1ac1a79161
commit 5e27e7e11c
2 changed files with 94 additions and 59 deletions
Download patch file Download diff file Expand all files Collapse all files

102
convert_hf_to_gguf.py
View file

@ -251,12 +251,7 @@ class Model:
return [(self.map_tensor_name(name), data_torch)]
def extra_f32_tensors(self, name: str, new_name: str, bid: int | None, n_dims: int) -> bool:
del name, new_name, bid, n_dims # unused
return False
def extra_f16_tensors(self, name: str, new_name: str, bid: int | None, n_dims: int) -> bool:
def tensor_force_quant(self, name: str, new_name: str, bid: int | None, n_dims: int) -> gguf.GGMLQuantizationType | bool:
del name, new_name, bid, n_dims # unused
return False
@ -285,55 +280,42 @@ class Model:
for new_name, data in ((n, d.squeeze().numpy()) for n, d in self.modify_tensors(data_torch, name, bid)):
data: np.ndarray # type hint
n_dims = len(data.shape)
data_dtype = data.dtype
data_qtype: gguf.GGMLQuantizationType | None = None
# when both are True, f32 should win
extra_f32 = self.extra_f32_tensors(name, new_name, bid, n_dims)
extra_f16 = self.extra_f16_tensors(name, new_name, bid, n_dims)
data_qtype: gguf.GGMLQuantizationType | bool = self.tensor_force_quant(name, new_name, bid, n_dims)
# Most of the codebase that takes in 1D tensors or norms only handles F32 tensors
# Conditions should closely match those in llama_model_quantize_internal in llama.cpp
extra_f32 = any(cond for cond in (
extra_f32,
n_dims == 1,
new_name.endswith("_norm.weight"),
))
# Some tensor types are always in float32
extra_f32 = extra_f32 or any(self.match_model_tensor_name(new_name, key, bid) for key in (
gguf.MODEL_TENSOR.FFN_GATE_INP,
gguf.MODEL_TENSOR.POS_EMBD,
gguf.MODEL_TENSOR.TOKEN_TYPES,
))
# if f16 desired, convert any float32 2-dim weight tensors to float16
extra_f16 = any(cond for cond in (
extra_f16,
(name.endswith(".weight") and n_dims >= 2),
))
if self.ftype != gguf.LlamaFileType.ALL_F32 and extra_f16 and not extra_f32:
if self.ftype == gguf.LlamaFileType.MOSTLY_BF16:
data = gguf.quantize_bf16(data)
assert data.dtype == np.uint16
data_qtype = gguf.GGMLQuantizationType.BF16
elif self.ftype == gguf.LlamaFileType.MOSTLY_Q8_0 and gguf.can_quantize_to_q8_0(data):
data = gguf.quantize_q8_0(data)
assert data.dtype == np.uint8
data_qtype = gguf.GGMLQuantizationType.Q8_0
else: # default to float16 for quantized tensors
if data_dtype != np.float16:
data = data.astype(np.float16)
data_qtype = gguf.GGMLQuantizationType.F16
if data_qtype is None: # by default, convert to float32
if data_dtype != np.float32:
data = data.astype(np.float32)
if n_dims <= 1 or new_name.endswith("_norm.weight"):
data_qtype = gguf.GGMLQuantizationType.F32
# Conditions should closely match those in llama_model_quantize_internal in llama.cpp
# Some tensor types are always in float32
if data_qtype is False and (
any(
self.match_model_tensor_name(new_name, key, bid)
for key in (
gguf.MODEL_TENSOR.FFN_GATE_INP,
gguf.MODEL_TENSOR.POS_EMBD,
gguf.MODEL_TENSOR.TOKEN_TYPES,
)
)
or not name.endswith(".weight")
):
data_qtype = gguf.GGMLQuantizationType.F32
if isinstance(data_qtype, bool):
data_qtype = gguf.LlamaFileTypeMap.get(self.ftype, gguf.GGMLQuantizationType.F32)
if data_qtype == gguf.GGMLQuantizationType.Q8_0:
if gguf.quants.Q8_0.can_quantize(data):
data = gguf.quants.Q8_0.quantize(data)
else: # fallback to f16
data_qtype = gguf.GGMLQuantizationType.F16
if data_qtype == gguf.GGMLQuantizationType.BF16:
data = gguf.quants.BF16.quantize(data)
if data_qtype == gguf.GGMLQuantizationType.F16:
data = data.astype(np.float16, copy=False)
if data_qtype == gguf.GGMLQuantizationType.F32:
data = data.astype(np.float32, copy=False)
shape = gguf.quant_shape_from_byte_shape(data.shape, data_qtype) if data.dtype == np.uint8 else data.shape
# reverse shape to make it similar to the internal ggml dimension order
@ -1765,7 +1747,7 @@ class DbrxModel(Model):
return [(new_name, data_torch)]
def extra_f16_tensors(self, name: str, new_name: str, bid: int | None, n_dims: int) -> bool:
def tensor_force_quant(self, name: str, new_name: str, bid: int | None, n_dims: int) -> gguf.GGMLQuantizationType | bool:
del name, new_name, bid # unused
return n_dims > 1
@ -2680,18 +2662,22 @@ class MambaModel(Model):
return [(new_name, data_torch)]
def extra_f32_tensors(self, name: str, new_name: str, bid: int | None, n_dims: int) -> bool:
del n_dims # unused
return bid is not None and new_name in (
self.format_tensor_name(n, bid, ".weight" if name.endswith(".weight") else "") for n in [
def tensor_force_quant(self, name: str, new_name: str, bid: int | None, n_dims: int) -> gguf.GGMLQuantizationType | bool:
if bid is not None and new_name in (
self.format_tensor_name(
n, bid, ".weight" if name.endswith(".weight") else ""
)
for n in [
gguf.MODEL_TENSOR.SSM_CONV1D,
gguf.MODEL_TENSOR.SSM_X,
gguf.MODEL_TENSOR.SSM_DT,
gguf.MODEL_TENSOR.SSM_A,
gguf.MODEL_TENSOR.SSM_D,
]
)
):
return gguf.GGMLQuantizationType.F32
return super().tensor_force_quant(name, new_name, bid, n_dims)
@Model.register("CohereForCausalLM")