File size: 22,895 Bytes
1e1a102
f5abf67
b6903c6
f5abf67
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
cc0a32f
f5abf67
 
 
 
ecf280c
 
f5abf67
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
6bb8448
f5abf67
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
cfb0d00
 
6bb8448
 
cfb0d00
 
 
 
 
 
 
f44cf4d
 
cfb0d00
 
 
 
 
 
 
 
f5abf67
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
9603085
f5abf67
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
ecf280c
f5abf67
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
8b96263
f5abf67
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
fbbbcab
f5abf67
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4c4b1d9
f5abf67
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
import spaces
import gradio as gr
gr.processing_utils._check_allowed = lambda path, allowed_paths: True
import io
import os
import time
import uuid
import traceback
import soundfile as sf
import torchaudio
import torch
from transformers import AutoModel, AutoProcessor, GenerationConfig, StoppingCriteria
from dataclasses import astuple
import sys


class MIMOStopper(StoppingCriteria):
    def __init__(self, stop_id: int) -> None:
        super().__init__()
        self.stop_id = stop_id

    def __call__(self, input_ids: torch.LongTensor, scores) -> bool:
        # Stop when last token of channel 0 is the stop token
        return input_ids[0, -1].item() == self.stop_id


class Inference:
    def __init__(self, model_path, codec_path=None, device='cuda'):
        self.device = device
        self.processor = AutoProcessor.from_pretrained(
            model_path,
            codec_path=codec_path if codec_path else "fnlp/MOSS-Speech",
            device=self.device,
            trust_remote_code=True,
        )
        self.model = AutoModel.from_pretrained(
            model_path, trust_remote_code=True
        ).to(self.device).eval()

    def forward(
        self,
        task: str,
        conversation_history_for_model: list, # Pass the entire conversation history formatted for the model
        temperature: float,
        top_p: float,
        repetition_penalty: float,
        max_new_tokens: int,
        min_new_tokens: int,
        top_k: int,
        system_prompt: str,
        decoder_audio_prompt_path: str = None
    ):
        # Prepare the conversation for the processor
        full_conversation = []
        if system_prompt:
            full_conversation.append({"role": "system", "content": system_prompt})
        
        # Add previous turns from the formatted history
        full_conversation.extend(conversation_history_for_model)

        output_modalities = []
        if task.endswith("speech_response"):
            output_modalities.append('audio')
        if task.endswith("text_response"):
            output_modalities.append('text')

        # This should always be exactly one modality based on task
        if len(output_modalities) != 1: 
            raise ValueError("Expected exactly one output modality based on task.")
        
        stopping_criteria = [
            MIMOStopper(self.processor.tokenizer.pad_token_id),
            MIMOStopper(
                self.processor.tokenizer.convert_tokens_to_ids("<|im_end|>"),
            ),
        ]
        
        generate_kwargs = {
            "temperature": temperature,
            "top_p": top_p,
            "repetition_penalty": repetition_penalty,
            "max_new_tokens": max_new_tokens,
            "min_new_tokens": min_new_tokens,
            "do_sample": True, # Always true for these parameters
            "use_cache": True,
            "top_k": top_k,
        }
        generation_config = GenerationConfig(**generate_kwargs)

        @spaces.GPU(duration = 120)
        def gen_spaces():
            inputs = self.processor([full_conversation], output_modalities)

            token_ids = self.model.generate(
                input_ids=inputs["input_ids"].to(self.device),
                attention_mask=inputs["attention_mask"].to(self.device),
                generation_config=generation_config,
                stopping_criteria=stopping_criteria
            )

            print(f"{token_ids.tolist()=}")

            results = self.processor.decode(
                token_ids.to(self.device),
                output_modalities,
                decoder_audio_prompt_path=decoder_audio_prompt_path
            )
            return results
        
        results = gen_spaces()
        
        # As per requirement, always one output modality, so take the first result
        response_obj = results[0] 

        text_out = None
        audio_out = None

        if output_modalities[0] == 'audio':
            audio_out = (response_obj.sampling_rate, response_obj.audio.squeeze(0).cpu().numpy()) if response_obj.audio is not None else None
        elif output_modalities[0] == 'text':
            text_out = response_obj.generated_text if response_obj.generated_text is not None else None

        # Clean up temporary user audio file if it was created (only temporary for processor)
        # if temp_user_audio_path and os.path.exists(temp_user_audio_path):
        #     os.remove(temp_user_audio_path)

        return text_out, audio_out


class MIMOInterface:
    def __init__(self, model_path):
        self.inference = Inference(model_path, codec_path="fnlp/MOSS-Speech-Codec")
        self.audio_dir = "chat_audio"
        os.makedirs(self.audio_dir, exist_ok=True)
        self.default_decoder_audio_prompt_path = "assets/prompt_cn.wav"


    # ---------- Helpers ----------

    def get_system_prompt_default(self, task):
        if task.endswith("speech_response"):
            return "You are a helpful voice assistant. Answer the user's questions with spoken responses."
        elif task.endswith("text_response"):
            return "You are a helpful assistant. Answer the user's questions with text."
        else:
            return "You are a helpful assistant."

    def _unique_wav_path(self, prefix: str) -> str:
        return os.path.join(self.audio_dir, f"{prefix}_{int(time.time()*1000)}_{uuid.uuid4().hex[:8]}.wav")

    def _save_audio_numpy(self, audio_np_tuple, prefix="audio") -> str:
        """
        audio_np_tuple: (sample_rate, np.ndarray)
        Returns local .wav path.
        """
        if audio_np_tuple is None:
            return ""
        
        sr, arr = audio_np_tuple
        if len(arr.shape) > 1:
            arr = arr[:, 0] # Ensure mono
        path = self._unique_wav_path(prefix)
        sf.write(path, arr, sr, format="WAV")
        return path

    def _delete_audio_files(self, file_paths: list):
        """Deletes a list of audio files."""
        for path in file_paths:
            if os.path.exists(path) and os.path.isfile(path):
                try:
                    os.remove(path)
                except Exception as e:
                    print(f"Error deleting audio file {path}: {e}")

    # ---------- Core inference + chat sync ----------

    def process_input(
        self,
        audio_input,
        text_input,
        mode,
        temperature,
        top_p,
        repetition_penalty,
        max_new_tokens,
        min_new_tokens,
        top_k,
        system_prompt,
        history_state_tuple, # (chatbot_messages, audio_file_paths_to_delete, conversation_for_model)
        decoder_audio_prompt # numpy tuple from gradio audio component
    ):
        chatbot_messages, audio_file_paths_to_delete, conversation_for_model = history_state_tuple
        # Keep a copy of the state before any changes in case of warning/error
        original_chatbot_messages = list(chatbot_messages)
        original_audio_file_paths_to_delete = list(audio_file_paths_to_delete)
        original_conversation_for_model = list(conversation_for_model)

        # new_chatbot_message = []
        try:
            # --- Handle Decoder Audio Prompt ---
            decoder_audio_prompt_path_for_model = None
            if decoder_audio_prompt:
                saved_decoder_audio_path = self._save_audio_numpy(decoder_audio_prompt, prefix="decoder_prompt")
                audio_file_paths_to_delete.append(saved_decoder_audio_path)
                decoder_audio_prompt_path_for_model = saved_decoder_audio_path
            else:
                decoder_audio_prompt_path_for_model = self.default_decoder_audio_prompt_path


            # --- Prepare User Input for Model and Display ---
            user_display_message_content = "" 
            user_audio_path_display = None 
            current_user_turn_for_model = None 

            if mode.startswith("speech_instruct"):
                if audio_input is None:
                    gr.Warning("Speech Input mode requires an audio input.")
                    return original_chatbot_messages[-1][1][0] if original_chatbot_messages else "", None, original_chatbot_messages, history_state_tuple # Return previous state
                else:
                    user_audio_path_display = self._save_audio_numpy(audio_input, prefix="user")
                    audio_file_paths_to_delete.append(user_audio_path_display)
                    user_display_message_content = "🎀 Voice message" # Consistent text for speech input

                    buffer = io.BytesIO()
                    sf.write(buffer, audio_input[1], audio_input[0], format="WAV")
                    buffer.seek(0)
                    current_user_turn_for_model = {"role": "user", "content": {'path': user_audio_path_display, 'type': 'audio/wav'}}
            else: # Text instruct modes
                txt = (text_input or "").strip()
                if not txt:
                    gr.Warning("Text Input mode requires a text input.")
                    return original_chatbot_messages[-1][1][0] if original_chatbot_messages else "", None, original_chatbot_messages, history_state_tuple # Return previous state
                else:
                    user_display_message_content = txt
                    current_user_turn_for_model = {"role": "user", "content": user_display_message_content}

            # Add user input to chatbot messages and model's conversation history
            # Always add a single entry for user turn in chatbot_messages
            if user_audio_path_display:
                # chatbot_messages.append([user_display_message_content, None]) 
                # new_chatbot_message.append([None, gr.Audio(user_audio_path_display, type='audio/wav')])
                chatbot_messages.append({'role': 'user', 'content': {'path': user_audio_path_display}})
            else:
                chatbot_messages.append({'role': 'user', 'content': user_display_message_content})
            
            if current_user_turn_for_model:
                conversation_for_model.append(current_user_turn_for_model)

            # --- Run Inference ---
            text_out, audio_out = self.inference.forward(
                task=mode,
                conversation_history_for_model=conversation_for_model,
                temperature=temperature,
                top_p=top_p,
                repetition_penalty=repetition_penalty,
                max_new_tokens=max_new_tokens,
                min_new_tokens=min_new_tokens,
                top_k=top_k,
                system_prompt=system_prompt,
                decoder_audio_prompt_path=decoder_audio_prompt_path_for_model
            )

            # --- Process Assistant Output for Display and Model History ---
            assistant_response_for_model_content = None # This will be string or dict for model history
            final_text_output_panel = None
            assistant_audio_output_panel = None
            
            # Assistant text for display/chatbot
            assistant_text_display = None
            assistant_audio_path_display = None

            if mode.endswith("speech_response"):
                if audio_out is None:
                    gr.Warning("Model failed to generate speech response.")
                    # Restore original history state if generation failed
                    return original_chatbot_messages[-1][1][0] if original_chatbot_messages else "", None, original_chatbot_messages, (original_chatbot_messages, original_audio_file_paths_to_delete, original_conversation_for_model)

                assistant_audio_output_panel = audio_out 
                saved_assistant_audio_path = self._save_audio_numpy(audio_out, prefix="assistant")
                audio_file_paths_to_delete.append(saved_assistant_audio_path)
                assistant_audio_path_display = saved_assistant_audio_path
                
                # Chatbot message for speech response mode
                # The text part is usually not needed, but can be a placeholder or empty
                # chatbot_messages.append(["πŸ”Š Generated speech.", None]) 
                # new_chatbot_message.append([None, gr.Audio(assistant_audio_path_display, type="filepath")])
                chatbot_messages.append({'role': 'assistant', 'content': {'path': assistant_audio_path_display}})
                assistant_response_for_model_content = {'path': saved_assistant_audio_path, 'type': 'filepath'}

            elif mode.endswith("text_response"):
                if text_out is None or str(text_out).strip() == "":
                    gr.Warning("Model failed to generate text response.")
                    # Restore original history state if generation failed
                    return original_chatbot_messages[-1][1][0] if original_chatbot_messages else "", None, original_chatbot_messages, (original_chatbot_messages, original_audio_file_paths_to_delete, original_conversation_for_model)

                final_text_output_panel = text_out
                assistant_text_display = text_out

                # Chatbot message for text response mode
                chatbot_messages.append({'role': 'assistant', 'content': assistant_text_display})
                assistant_response_for_model_content = text_out

            # Add assistant's actual response to the conversation for the next model turn
            if assistant_response_for_model_content:
                conversation_for_model.append({"role": "assistant", "content": assistant_response_for_model_content})
            
            # Return updated history state tuple
            new_history_state_tuple = (chatbot_messages, audio_file_paths_to_delete, conversation_for_model)
            # Return panel outputs + chat + state
            return final_text_output_panel, assistant_audio_output_panel, chatbot_messages, new_history_state_tuple

        except Exception as e:
            traceback.print_exc()
            err = f"Error: {str(e)}"
            gr.Error(f"An unexpected error occurred: {err}")
            # Restore original history state on any unhandled exception
            return original_chatbot_messages[-1][0] if original_chatbot_messages else "", None, original_chatbot_messages, (original_chatbot_messages, original_audio_file_paths_to_delete, original_conversation_for_model)

    def _submit_with_clear(
        self, audio_in, text_in, mode, temperature, top_p, repetition_penalty, max_new_tokens, min_new_tokens, top_k,
        system_prompt, history_state, decoder_audio_prompt, clear_on_submit
    ):
        if clear_on_submit:
            _, audio_files, _ = history_state
            self._delete_audio_files(audio_files)
            history_state = ([], [], [])
        return self.process_input(
            audio_in, text_in, mode, temperature, top_p, repetition_penalty,
            max_new_tokens, min_new_tokens, top_k, system_prompt,
            history_state, decoder_audio_prompt
        )
        
    # ---------- UI factory ----------

    def create_interface(self):
        theme = gr.themes.Soft()

        with gr.Blocks(theme=theme) as demo:
            gr.HTML(
                """
                <div class="main-header">
                    <h1>🎀 MOSS-Speech Demo</h1>
                </div>
                """
            )


            mode = gr.Radio(
                [
                    ("Speech In β†’ Speech Out", "speech_instruct_speech_response"),
                    ("Speech In β†’ Text Out", "speech_instruct_text_response"),
                    ("Text In β†’ Speech Out", "text_instruct_speech_response"),
                    ("Text In β†’ Text Out", "text_instruct_text_response"),
                ],
                label="🎯 Interaction Mode",
                value="speech_instruct_speech_response",
                container=True,
                scale=1,
            )

            system_prompt = gr.Textbox(
                label="πŸ€– System Prompt",
                value=self.get_system_prompt_default("speech_instruct_speech_response"),
                lines=2,
                container=True,
                scale=1,
            )

            with gr.Accordion("βš™οΈ Generation Parameters", open=False, elem_classes="param-accordion"):
                with gr.Row():
                    temperature = gr.Slider(0.1, 2.0, value=0.6, step=0.1, label="🌑️ Temperature", info="Higher = more random")
                    top_p = gr.Slider(0.1, 1.0, value=0.95, step=0.05, label="🎯 Top-p", info="Nucleus sampling")
                    top_k = gr.Slider(1, 100, value=20, step=1, label="πŸ” Top-k", info="Candidate tokens")

                with gr.Row():
                    repetition_penalty = gr.Slider(1.0, 2.0, value=1.1, step=0.1, label="πŸ”„ Repetition Penalty", info="Discourage repeats")
                    max_new_tokens = gr.Slider(1, 2000, value=500, step=1, label="πŸ“ Max New Tokens", info="Upper bound")
                    min_new_tokens = gr.Slider(0, 100, value=0, step=1, label="πŸ“ Min New Tokens", info="Lower bound")
                    
                decoder_audio_prompt = gr.Audio(type="numpy", label="πŸŽ™οΈ Decoder Audio Prompt (Optional)", visible=True)

            with gr.Row():
                with gr.Column(scale=1, elem_classes="input-section"):
                    gr.Markdown("### πŸ“₯ Input")

                    audio_input = gr.Audio(type="numpy", label="πŸŽ™οΈ Speech Input", visible=True)

                    text_input = gr.Textbox(
                        label="🧾 Text Input",
                        placeholder="Type your question here…",
                        lines=3,
                        info="Enter text to query the assistant",
                        visible=False,
                    )

                with gr.Column(scale=1, elem_classes="output-section"):
                    gr.Markdown("### πŸ“€ Output")

                    text_output = gr.Textbox(
                        label="πŸ“„ Text Output",
                        lines=8,
                        interactive=False,
                        info="Model-generated text response",
                        visible=False,
                    )

                    audio_output = gr.Audio(label="πŸ”Š Speech Output", visible=True, autoplay=True)
                    
            with gr.Row():
                submit_btn = gr.Button("πŸš€ Submit", variant="primary", elem_classes="btn-primary")
                clear_history_btn = gr.Button("πŸ—‘οΈ Clear All History", variant="secondary", elem_classes="btn-secondary")
                
            with gr.Row():
                clear_history_on_mode_change_checkbox = gr.Checkbox(
                    label="Clear history on mode change", value=True, interactive=True
                )
                clear_history_on_submit_checkbox = gr.Checkbox(
                    label="Clear history on each submit", value=False, interactive=True
                )

            # history_state will now be a tuple: (chatbot_messages, audio_file_paths_to_delete, conversation_for_model)
            history_state = gr.State(([], [], [])) 
            chatbot = gr.Chatbot(
                elem_id="chatbot",
                bubble_full_width=True,
                type="messages", # Keep commented to allow [text, audio] in chatbot
                scale=1,
                label="πŸ’¬ Chat History",
                show_copy_button=True
            )


            # ---------- Event handlers ----------

            submit_btn.click(
                fn=self._submit_with_clear,
                inputs=[
                    audio_input,
                    text_input,
                    mode,
                    temperature,
                    top_p,
                    repetition_penalty,
                    max_new_tokens,
                    min_new_tokens,
                    top_k,
                    system_prompt,
                    history_state, # Pass the current Gradio state tuple
                    decoder_audio_prompt,
                    clear_history_on_submit_checkbox
                ],
                outputs=[text_output, audio_output, chatbot, history_state],
            )

            def _hard_clear(current_history_state_tuple):
                _, audio_files, _ = current_history_state_tuple
                self._delete_audio_files(audio_files)
                gr.Info("Conversation history and associated audio files cleared.")
                return "", None, [], ([], [], [])

            clear_history_btn.click(
                fn=_hard_clear,
                inputs=[history_state],
                outputs=[text_output, audio_output, chatbot, history_state],
            )

            def update_interface_visibility(selected_mode):
                if selected_mode.startswith("speech_instruct"):
                    return gr.update(visible=True), gr.update(visible=False)
                else:
                    return gr.update(visible=False), gr.update(visible=True)

            def update_output_visibility(selected_mode):
                if selected_mode.endswith("speech_response"):
                    return gr.update(visible=False), gr.update(visible=True)
                else:
                    return gr.update(visible=True), gr.update(visible=False)

            def _on_mode_change(task, clear_history_on_mode_change, current_history_state_tuple):
                if clear_history_on_mode_change:
                    _, audio_files_to_delete, _ = current_history_state_tuple
                    self._delete_audio_files(audio_files_to_delete)
                    gr.Info("Interaction mode changed. History cleared.")
                    return self.get_system_prompt_default(task), [], ([], [], [])
                else:
                    gr.Info("Interaction mode changed. History preserved.")
                    # Keep existing chatbot messages and state
                    chatbot_messages, audio_files, conv_state = current_history_state_tuple
                    return self.get_system_prompt_default(task), chatbot_messages, (chatbot_messages, audio_files, conv_state)

            mode.change(
                fn=_on_mode_change,
                inputs=[mode, clear_history_on_mode_change_checkbox, history_state],
                outputs=[system_prompt, chatbot, history_state],
            )
            mode.change(
                fn=update_interface_visibility,
                inputs=[mode],
                outputs=[audio_input, text_input],
            )
            mode.change(
                fn=update_output_visibility,
                inputs=[mode],
                outputs=[text_output, audio_output],
            )

        return demo

if __name__ == "__main__":
    model_path = "fnlp/MOSS-Speech"
    
    interface = MIMOInterface(model_path)
    demo = interface.create_interface()
    demo.launch()