* feat: add domyn-edge model * Generate tokenizer hash for conversion + fix tests * Address changes according to code-review skill * catch ValueError in convert_hf_to_gguf_update.py * renegerate fusion test baseline + fix model naming style
54 lines
2.7 KiB
Python
54 lines
2.7 KiB
Python
from .base import TextModel, ModelBase
|
|
import gguf
|
|
|
|
|
|
@ModelBase.register("DomynEdgeForCausalLM")
|
|
@ModelBase.example("domyn/Domyn-Edge-Teacher-IF")
|
|
class DomynEdgeModel(TextModel):
|
|
model_arch = gguf.MODEL_ARCH.DOMYN_EDGE
|
|
|
|
def set_gguf_parameters(self):
|
|
super().set_gguf_parameters()
|
|
hparams = self.hparams
|
|
self.gguf_writer.add_vocab_size(hparams["vocab_size"])
|
|
|
|
position_layers = hparams["position_layers"] # list of "rope"/"nope" strings per layer
|
|
head_sizes = hparams["head_sizes"] # list of int head dims per layer
|
|
|
|
# Per-layer rope/SWA pattern: True="rope", False="nope"
|
|
rope_pattern = [v == "rope" for v in position_layers]
|
|
self.gguf_writer.add_rope_pattern(rope_pattern)
|
|
self.gguf_writer.add_sliding_window_pattern(rope_pattern) # SWA <-> RoPE in domyn-edge
|
|
|
|
# window_sizes entries are flash-attn style [left, right] (current token excluded; -1 = full attention),
|
|
# while llama.cpp's n_swa counts the current token, hence the +1 (matches HF's sliding_window=1025)
|
|
swa_window = hparams.get("sliding_window")
|
|
if swa_window is None:
|
|
window_sizes = hparams.get("window_sizes") # list of [size, offset] pairs per layer
|
|
if window_sizes is None:
|
|
raise KeyError("Neither `sliding_window` nor `window_sizes` found among domyn edge hyperparameters")
|
|
left = next((ws[0] for v, ws in zip(position_layers, window_sizes) if v == "rope"), None)
|
|
swa_window = left + 1 if left is not None and left > 0 else None
|
|
if swa_window is None:
|
|
raise ValueError("No sliding window size found among domyn edge hyperparameters")
|
|
self.gguf_writer.add_sliding_window(swa_window)
|
|
|
|
# Head dims differ between full-attention ("nope") and SWA ("rope") layers
|
|
full_head_dims = {h for v, h in zip(position_layers, head_sizes) if v == "nope"}
|
|
swa_head_dims = {h for v, h in zip(position_layers, head_sizes) if v == "rope"}
|
|
|
|
if len(full_head_dims) != 1:
|
|
raise ValueError(f"Expected one head dimension for full attention layers, found {len(full_head_dims)}")
|
|
full_head_dim = full_head_dims.pop()
|
|
|
|
self.gguf_writer.add_key_length(full_head_dim)
|
|
self.gguf_writer.add_value_length(full_head_dim)
|
|
|
|
if len(swa_head_dims) != 1:
|
|
raise ValueError(f"Expected one head dimension for SWA attention layers, found {len(swa_head_dims)}")
|
|
swa_head_dim = swa_head_dims.pop()
|
|
|
|
self.gguf_writer.add_key_length_swa(swa_head_dim)
|
|
self.gguf_writer.add_value_length_swa(swa_head_dim)
|
|
self.gguf_writer.add_rope_dimension_count_swa(swa_head_dim)
|