Upload model.py with huggingface_hub
Browse files
model.py
ADDED
|
@@ -0,0 +1,80 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import mlx.core as mx
|
| 2 |
+
import mlx.nn as nn
|
| 3 |
+
from mlx_lm.models.base import BaseModelArgs
|
| 4 |
+
from mlx_lm.utils import load
|
| 5 |
+
|
| 6 |
+
from subsampling import ConvSubsampling
|
| 7 |
+
from modules import ConformerLayer
|
| 8 |
+
from attention import RelPositionalEncoding
|
| 9 |
+
|
| 10 |
+
class FastConformerEncoder(nn.Module):
|
| 11 |
+
def __init__(
|
| 12 |
+
self,
|
| 13 |
+
feat_in: int = 80,
|
| 14 |
+
n_layers: int = 18,
|
| 15 |
+
d_model: int = 512,
|
| 16 |
+
ff_expansion_factor: int = 4,
|
| 17 |
+
n_heads: int = 8,
|
| 18 |
+
conv_kernel_size: int = 9,
|
| 19 |
+
dropout: float = 0.1,
|
| 20 |
+
):
|
| 21 |
+
super().__init__()
|
| 22 |
+
self.d_model = d_model
|
| 23 |
+
# FastConformer uses 8x downsampling via striding
|
| 24 |
+
self.pre_encode = ConvSubsampling(
|
| 25 |
+
subsampling='dw_striding',
|
| 26 |
+
subsampling_factor=8,
|
| 27 |
+
feat_in=feat_in,
|
| 28 |
+
feat_out=d_model,
|
| 29 |
+
conv_channels=256
|
| 30 |
+
)
|
| 31 |
+
|
| 32 |
+
self.pos_enc = RelPositionalEncoding(d_model=d_model, max_len=5000)
|
| 33 |
+
|
| 34 |
+
d_ff = d_model * ff_expansion_factor
|
| 35 |
+
self.layers = [
|
| 36 |
+
ConformerLayer(
|
| 37 |
+
d_model=d_model,
|
| 38 |
+
d_ff=d_ff,
|
| 39 |
+
n_heads=n_heads,
|
| 40 |
+
conv_kernel_size=conv_kernel_size,
|
| 41 |
+
dropout=dropout
|
| 42 |
+
) for _ in range(n_layers)
|
| 43 |
+
]
|
| 44 |
+
|
| 45 |
+
def __call__(self, x, lengths=None):
|
| 46 |
+
# x: (batch, time, feat_in)
|
| 47 |
+
if lengths is None:
|
| 48 |
+
lengths = mx.array([x.shape[1]] * x.shape[0])
|
| 49 |
+
|
| 50 |
+
x, lengths = self.pre_encode(x, lengths)
|
| 51 |
+
|
| 52 |
+
# Add positional encoding
|
| 53 |
+
x, pos_emb = self.pos_enc(x)
|
| 54 |
+
|
| 55 |
+
for layer in self.layers:
|
| 56 |
+
x = layer(x, pos_emb=pos_emb)
|
| 57 |
+
|
| 58 |
+
return x, lengths
|
| 59 |
+
|
| 60 |
+
class CanaryModel(nn.Module):
|
| 61 |
+
"""
|
| 62 |
+
Hybrid ASR-LLM Model connecting FastConformer to Qwen via a linear projection.
|
| 63 |
+
"""
|
| 64 |
+
def __init__(self, encoder: FastConformerEncoder, llm_dim: int):
|
| 65 |
+
super().__init__()
|
| 66 |
+
self.encoder = encoder
|
| 67 |
+
|
| 68 |
+
# Audio to Text projection (LoRA or standard linear depending on NeMo config)
|
| 69 |
+
self.audio_encoder_proj = nn.Linear(1024, llm_dim)
|
| 70 |
+
|
| 71 |
+
def encode_audio(self, audio_features, lengths=None):
|
| 72 |
+
"""
|
| 73 |
+
Passes audio features through the encoder and projects them to the LLM embedding space.
|
| 74 |
+
"""
|
| 75 |
+
encoded_audio, lengths = self.encoder(audio_features, lengths)
|
| 76 |
+
audio_embeds = self.audio_encoder_proj(encoded_audio)
|
| 77 |
+
return audio_embeds, lengths
|
| 78 |
+
|
| 79 |
+
# Note: Text generation logic will be handled externally by feeding
|
| 80 |
+
# the audio_embeds directly into the mlx_lm Qwen decoder prompt embedding!
|