removes speaker encoder ckpt (metavoiceio#51)

* feat: remove speaker encoder cpkt

* fix: missing attr

---------

Co-authored-by: sid <sid@themetavoice.xyz>

fam/llm/sample.py

@@ -561,8 +561,10 @@ def sample_utterance(
     )[0]
 
 
-def build_models(config_first_stage, config_second_stage, device, use_kv_cache):
-    smodel = SpeakerEncoder(device=device, eval=True, verbose=False)
+def build_models(config_first_stage, config_second_stage, model_dir, device, use_kv_cache):
+    smodel = SpeakerEncoder(
+        weights_fpath=os.path.join(model_dir, "speaker_encoder.pt"), device=device, eval=True, verbose=False
+    )
     data_adapter = FlattenedInterleavedEncodec2Codebook(end_of_audio_token=1024)
     llm_first_stage = Model(
         config_first_stage,
@@ -690,7 +692,11 @@ class SamplingControllerConfig:
 
     # define models
     smodel, llm_first_stage, llm_second_stage = build_models(
-        config_first_stage, config_second_stage, sampling_config.device, sampling_config.use_kv_cache
+        config_first_stage,
+        config_second_stage,
+        model_dir=model_dir,
+        device=sampling_config.device,
+        use_kv_cache=sampling_config.use_kv_cache
     )
 
     print(f"Synthesising utterance...")

fam/llm/serving.py

@@ -56,6 +56,9 @@ class ServingConfig:
     enhancer: Optional[Literal["df"]] = "df"
     """Enhancer to use for post-processing."""
 
+    compile: bool = False
+    """Whether to compile the model using PyTorch 2.0."""
+
     port: int = 58003
 
 
@@ -181,7 +184,9 @@ def _convert_audiodata_to_wav_path(audiodata, wav_tmp):
         **common_config,
     )
 
-    spkemb, llm_stg1, llm_stg2 = build_models(config1, config2, device=device, use_kv_cache="flash_decoding")
+    spkemb, llm_stg1, llm_stg2 = build_models(
+        config1, config2, model_dir=model_dir, device=device, use_kv_cache="flash_decoding"
+    )
     GlobalState.spkemb_model = spkemb
     GlobalState.first_stage_model = llm_stg1
     GlobalState.second_stage_model = llm_stg2

fam/quantiser/audio/speaker_encoder/model.py

@@ -9,8 +9,6 @@
 
 from fam.quantiser.audio.speaker_encoder import audio
 
-DEFAULT_SPKENC_CKPT_PATH = os.path.join(os.path.dirname(os.path.abspath(__file__)), "ckpt/ckpt.pt")
-
 mel_window_step = 10
 mel_n_channels = 40
 sampling_rate = 16000
@@ -43,14 +41,10 @@ def __init__(
         self.device = device
 
         start = timer()
-        if eval and weights_fpath is None:
-            weights_fpath = DEFAULT_SPKENC_CKPT_PATH
-
-        if weights_fpath is not None:
-            checkpoint = torch.load(weights_fpath, map_location="cpu")
 
-            self.load_state_dict(checkpoint["model_state"], strict=False)
-            self.to(device)
+        checkpoint = torch.load(weights_fpath, map_location="cpu")
+        self.load_state_dict(checkpoint["model_state"], strict=False)
+        self.to(device)
 
         if eval:
             self.eval()