wehub-resource-sync
106 lines
3.4 KiB
Python
106 lines
3.4 KiB
Python
"""
|
|
This file specifies how MLC's EAGLE parameter maps from other formats, for example HuggingFace
|
|
PyTorch, HuggingFace safetensors.
|
|
"""
|
|
|
|
import functools
|
|
|
|
import numpy as np
|
|
|
|
from mlc_llm.loader import ExternMapping
|
|
from mlc_llm.loader.standard_loader import make_standard_hf_loader
|
|
from mlc_llm.quantization import Quantization, make_awq_quant
|
|
|
|
from .eagle_model import EagleConfig, EagleForCausalLM
|
|
|
|
awq_quant = make_awq_quant(EagleForCausalLM)
|
|
|
|
|
|
huggingface = make_standard_hf_loader(
|
|
model_cls=EagleForCausalLM,
|
|
layer_prefix="layers",
|
|
add_unused=["rotary_emb.inv_freq"],
|
|
)
|
|
|
|
|
|
def awq(model_config: EagleConfig, quantization: Quantization) -> ExternMapping:
|
|
"""Returns a parameter mapping that maps from the names of MLC LLM parameters to
|
|
the names of AWQ parameters.
|
|
Parameters
|
|
----------
|
|
model_config : EagleConfig
|
|
The configuration of the Eagle model.
|
|
|
|
quantization : Quantization
|
|
The quantization configuration.
|
|
|
|
Returns
|
|
-------
|
|
param_map : ExternMapping
|
|
The parameter mapping from MLC to AWQ.
|
|
"""
|
|
model, _ = awq_quant(model_config, quantization)
|
|
_, _named_params, _ = model.export_tvm(
|
|
spec=model.get_default_spec(),
|
|
allow_extern=True,
|
|
)
|
|
named_parameters = dict(_named_params)
|
|
|
|
mapping = ExternMapping()
|
|
|
|
for i in range(model_config.num_hidden_layers):
|
|
# Add QKV in self attention
|
|
attn = f"layers.{i}.self_attn"
|
|
for quantize_suffix in ["qweight", "qzeros", "scales"]:
|
|
mlc_name = f"{attn}.qkv_proj.{quantize_suffix}"
|
|
assert mlc_name in named_parameters
|
|
mlc_param = named_parameters[mlc_name]
|
|
mapping.add_mapping(
|
|
mlc_name,
|
|
[
|
|
f"{attn}.q_proj.{quantize_suffix}",
|
|
f"{attn}.k_proj.{quantize_suffix}",
|
|
f"{attn}.v_proj.{quantize_suffix}",
|
|
],
|
|
functools.partial(
|
|
lambda q, k, v, dtype: np.concatenate(
|
|
[q, k, v],
|
|
axis=1, # AWQ GEMM would transpose the weight
|
|
).astype(dtype),
|
|
dtype=mlc_param.dtype,
|
|
),
|
|
)
|
|
|
|
# Concat gate and up in MLP
|
|
mlp = f"layers.{i}.mlp"
|
|
for quantize_suffix in ["qweight", "qzeros", "scales"]:
|
|
mlc_name = f"{mlp}.gate_up_proj.{quantize_suffix}"
|
|
assert mlc_name in named_parameters
|
|
mlc_param = named_parameters[mlc_name]
|
|
mapping.add_mapping(
|
|
mlc_name,
|
|
[
|
|
f"{mlp}.gate_proj.{quantize_suffix}",
|
|
f"{mlp}.up_proj.{quantize_suffix}",
|
|
],
|
|
functools.partial(
|
|
lambda gate, up, dtype: np.concatenate(
|
|
[gate, up],
|
|
axis=1, # AWQ GEMM would transpose the weight
|
|
).astype(dtype),
|
|
dtype=mlc_param.dtype,
|
|
),
|
|
)
|
|
|
|
# inv_freq is not used in the model
|
|
mapping.add_unused(f"{attn}.rotary_emb.inv_freq")
|
|
|
|
for mlc_name, mlc_param in named_parameters.items():
|
|
if mlc_name not in mapping.param_map:
|
|
mapping.add_mapping(
|
|
mlc_name,
|
|
[mlc_name],
|
|
functools.partial(lambda x, dtype: x.astype(dtype), dtype=mlc_param.dtype),
|
|
)
|
|
return mapping
|