Chatterbox-Finnish-ONNX / scripts /browser_pipeline_sim.py

Add scripts/browser_pipeline_sim.py

6264e00 verified 8 months ago

12.7 kB

"""
browser_pipeline_sim.py
Python simulation of the browser worker.ts logic.
Mirrors the EXACT token assembly, CFG loop, sampling, and decoding
that the browser worker does — making it trivial to diff against worker.ts.
If this produces correct audio, worker.ts with the same logic will too.
If there's a bug in the browser, you can find it by diffing this script against worker.ts.
Usage:
    cd /workspaces/work
    conda run -n chatterbox-onnx python browser_pipeline_sim.py [text] [ref_audio]
Outputs:
    _cmp/browser_sim_output.wav
"""

import os, sys, time, json, struct, math
import numpy as np
import onnxruntime as ort
import soundfile as sf
import librosa
import requests
from pathlib import Path
from huggingface_hub import hf_hub_download

sys.path.insert(0, "Chatterbox-Finnish")

# ── Config — mirrors browser worker constants ─────────────────────────────────
TEXT       = sys.argv[1] if len(sys.argv) > 1 else \
             "Tervetuloa kokeilemaan hienoviritettyä suomenkielistä Chatterbox-puhesynteesiä."
REF_AUDIO  = sys.argv[2] if len(sys.argv) > 2 else \
             "Chatterbox-Finnish/samples/reference_finnish.wav"

HF_BASE    = "onnx-community/chatterbox-multilingual-ONNX"
HF_FI      = "RASMUS/Chatterbox-Finnish-ONNX"
CACHE_DIR  = "_onnx_cache"
OUT_DIR    = Path("_cmp"); OUT_DIR.mkdir(exist_ok=True)

# Worker constants (from worker.ts)
SOT            = 255    # [START] token
EOT            = 0      # [STOP] token
START_SPEECH   = 6561   # BOS speech token
STOP_SPEECH    = 6562   # EOS speech token
CFG_WEIGHT     = 0.5
REP_PENALTY    = 1.2
TEMPERATURE    = 0.8
EXAGGERATION   = 0.6
MIN_SPEECH_TOKENS = 40
MAX_STEPS      = 1000
SAMPLE_RATE    = 24000

GROQ_KEY = os.environ.get("GROQ_API_KEY", "")

# ── puncNorm — mirrors worker.ts puncNorm() ───────────────────────────────────
def punc_norm(text: str) -> str:
    """Mirrors puncNorm() in worker.ts"""
    import re
    t = text.strip()
    t = re.sub(r'\s+', ' ', t)                         # normalize whitespace
    t = t[0].upper() + t[1:] if t else t               # capitalize first letter
    t = t.replace(' .', '.').replace(' ,', ',')        # remove space before punct
    t = t.replace(' ?', '?').replace(' !', '!')
    if t and t[-1] not in '.!?…':                      # ensure ending punctuation
        t += '.'
    return t


# ── EnTokenizer (same as Chatterbox-Finnish/src/chatterbox_/models/tokenizers/tokenizer.py)
def load_tokenizer(tokenizer_path: str):
    from src.chatterbox_.models.tokenizers.tokenizer import EnTokenizer
    return EnTokenizer(tokenizer_path)


# ── puncNorm + tokenize → mirrors worker.ts tokenize() ───────────────────────
def tokenize(tok, text: str) -> list[int]:
    """
    Mirrors browser worker tokenize():
    1. puncNorm
    2. replace spaces with [SPACE] token
    3. encode grapheme by grapheme
    4. wrap with [SOT, ..., EOT]
    """
    normed = punc_norm(text)
    ids = tok.encode(normed)
    return [SOT] + ids + [EOT]


# ── Download helpers ───────────────────────────────────────────────────────────
def dl(repo_id, filename):
    return hf_hub_download(repo_id=repo_id, filename=filename,
                           local_dir=CACHE_DIR, local_dir_use_symlinks=False)


# ── KV cache helpers ───────────────────────────────────────────────────────────
def empty_kv(n_layers=30, kv_dtype=np.float32):
    return [(np.zeros((1, 16, 0, 64), dtype=kv_dtype),
             np.zeros((1, 16, 0, 64), dtype=kv_dtype))
            for _ in range(n_layers)]


def make_kv_feeds(kv_cache):
    feeds = {}
    for i, (k, v) in enumerate(kv_cache):
        feeds[f"past_key_values.{i}.key"]   = k
        feeds[f"past_key_values.{i}.value"] = v
    return feeds


# ── Sampling helpers ───────────────────────────────────────────────────────────
def apply_rep_penalty(logits, generated_set, penalty):
    logits = logits.copy()
    for tok in generated_set:
        logits[tok] = logits[tok] / penalty if logits[tok] > 0 else logits[tok] * penalty
    return logits


def apply_min_p(logits, p=0.05):
    """Mirrors applyMinP() in worker.ts"""
    logits = logits.copy()
    probs = np.exp(logits - logits.max())
    probs /= probs.sum()
    threshold = probs.max() * p
    logits[probs < threshold] = -1e9
    return logits


def sample_with_temperature(logits, temperature):
    """Mirrors sampleWithTemperature() in worker.ts"""
    logits = (logits / temperature).astype(np.float64)
    logits -= logits.max()
    probs = np.exp(logits)
    probs /= probs.sum()
    return int(np.random.choice(len(probs), p=probs))


# ── Main pipeline ─────────────────────────────────────────────────────────────
def main():
    providers = ["CUDAExecutionProvider", "CPUExecutionProvider"]
    opts = ort.SessionOptions()
    opts.graph_optimization_level = ort.GraphOptimizationLevel.ORT_ENABLE_ALL

    # ── Load sessions ──
    print("Loading ONNX sessions...")
    sess_se = ort.InferenceSession(dl(HF_BASE, "onnx/speech_encoder.onnx"),  opts, providers)
    sess_et = ort.InferenceSession(dl(HF_BASE, "onnx/embed_tokens.onnx"),    opts, providers)
    sess_lm = ort.InferenceSession(dl(HF_FI,   "onnx/language_model.onnx"),  opts, providers)
    sess_cd = ort.InferenceSession(dl(HF_BASE, "onnx/conditional_decoder.onnx"), opts, providers)

    # ── Precomputed Finnish conditioning ──
    cond_emb_raw = open(dl(HF_FI, "onnx/finnish_cond_emb.bin"), "rb").read()
    cond_emb = np.frombuffer(cond_emb_raw, dtype=np.float32).reshape(1, 34, 1024)
    print(f"  cond_emb: {cond_emb.shape}")

    # ── KV dtype from model ──
    kv_input_name = "past_key_values.0.key"
    kv_dtype_str = next(
        (inp.type for inp in sess_lm.get_inputs() if inp.name == kv_input_name), "tensor(float)"
    )
    kv_dtype = np.float16 if "float16" in kv_dtype_str else np.float32
    print(f"  KV cache dtype: {kv_dtype}")

    # ── Step 1: Reference audio → speaker embeddings ──
    print(f"\nStep 1: speech_encoder  ({REF_AUDIO})")
    ref_audio, ref_sr = librosa.load(REF_AUDIO, sr=None)
    ref_16k = librosa.resample(ref_audio, orig_sr=ref_sr, target_sr=16000).astype(np.float32)
    ref_input = ref_16k[np.newaxis, :]  # [1, T]

    se_outs = sess_se.run(None, {"audio": ref_input})
    speaker_emb   = se_outs[0]  # [1, 256]
    prompt_tokens = se_outs[1]  # [1, N]
    print(f"  speaker_emb: {speaker_emb.shape}")
    print(f"  prompt_tokens: {prompt_tokens.shape}")
    # Note: se_outs[2] is cond_emb from base model — we ignore it, use precomputed Finnish version

    # ── Step 2: Tokenize text (mirrors worker.ts tokenize()) ──
    print(f"\nStep 2: tokenize  '{TEXT[:60]}...'")
    tok = load_tokenizer("Chatterbox-Finnish/pretrained_models/tokenizer.json")
    token_ids = tokenize(tok, TEXT)
    print(f"  token_ids ({len(token_ids)}): {token_ids[:6]}...{token_ids[-3:]}")
    text_ids_np = np.array([token_ids], dtype=np.int64)  # [1, T]

    # ── Step 3: Embed text tokens ──
    print(f"\nStep 3: embed_tokens")
    text_embeds = sess_et.run(None, {"input_ids": text_ids_np})[0]  # [1, T, 1024]
    print(f"  text_embeds: {text_embeds.shape}")

    # Embed BOS speech token
    bos_ids = np.array([[START_SPEECH]], dtype=np.int64)
    bos_emb = sess_et.run(None, {"input_ids": bos_ids})[0]  # [1, 1, 1024]

    # ── Step 4: Build prefill ──
    # Matches PyTorch: inputs_embeds = cat([cond_emb, text_emb, bos_emb])
    prefill_cond   = np.concatenate([cond_emb, text_embeds, bos_emb], axis=1)
    zeros_text     = np.zeros_like(text_embeds)
    prefill_uncond = np.concatenate([cond_emb, zeros_text, bos_emb], axis=1)

    mask_cond   = np.ones((1, prefill_cond.shape[1]),   dtype=np.int64)
    mask_uncond = np.ones((1, prefill_uncond.shape[1]), dtype=np.int64)

    kv_empty_layer = np.zeros((1, 16, 0, 64), dtype=kv_dtype)
    kv_cond   = [(kv_empty_layer.copy(), kv_empty_layer.copy()) for _ in range(30)]
    kv_uncond = [(kv_empty_layer.copy(), kv_empty_layer.copy()) for _ in range(30)]

    def lm_step(embeds, mask, kv):
        feeds = {"inputs_embeds": embeds, "attention_mask": mask}
        feeds.update(make_kv_feeds(kv))
        outs = sess_lm.run(None, feeds)
        logits = outs[0]  # [1, seq, vocab]
        new_kv = [(outs[1 + i*2], outs[1 + i*2 + 1]) for i in range(30)]
        return logits, new_kv

    # ── Step 5: Prefill both streams ──
    print(f"\nStep 4: prefill  ({prefill_cond.shape[1]} tokens)")
    t0 = time.time()
    logits_c,  kv_cond   = lm_step(prefill_cond,   mask_cond,   kv_cond)
    logits_uc, kv_uncond = lm_step(prefill_uncond, mask_uncond, kv_uncond)
    print(f"  prefill done ({time.time()-t0:.1f}s)")

    # ── Step 6: Autoregressive generation ──
    print(f"\nStep 5: generate (max {MAX_STEPS} steps)")
    generated = [START_SPEECH]
    speech_tokens = []
    t0 = time.time()

    for step in range(MAX_STEPS):
        last_ids = np.array([[generated[-1]]], dtype=np.int64)
        last_emb = sess_et.run(None, {"input_ids": last_ids})[0]  # [1, 1, 1024]

        seq_len_c  = kv_cond[0][0].shape[2] + 1
        seq_len_uc = kv_uncond[0][0].shape[2] + 1
        mask_c  = np.ones((1, seq_len_c),  dtype=np.int64)
        mask_uc = np.ones((1, seq_len_uc), dtype=np.int64)

        logits_c,  kv_cond   = lm_step(last_emb, mask_c,  kv_cond)
        logits_uc, kv_uncond = lm_step(last_emb, mask_uc, kv_uncond)

        # CFG: final = cond + cfg_weight * (cond - uncond)
        lc  = logits_c[0, -1].astype(np.float32)
        luc = logits_uc[0, -1].astype(np.float32)
        final_logits = lc + CFG_WEIGHT * (lc - luc)

        # Apply rep penalty + min_p + temperature sample
        final_logits = apply_rep_penalty(final_logits, set(generated), REP_PENALTY)
        final_logits = apply_min_p(final_logits, p=0.05)
        token = sample_with_temperature(final_logits, TEMPERATURE)

        if token == STOP_SPEECH and len(speech_tokens) >= MIN_SPEECH_TOKENS:
            print(f"  EOS at step {step} ({len(speech_tokens)} speech tokens)")
            break

        generated.append(token)
        if token < START_SPEECH:
            speech_tokens.append(token)

        if (step + 1) % 100 == 0:
            elapsed = time.time() - t0
            rate = (step + 1) / elapsed
            print(f"  step {step+1:4d}: {len(speech_tokens):3d} speech tokens  ({rate:.1f} tok/s)")

    elapsed = time.time() - t0
    print(f"  generation done: {len(speech_tokens)} speech tokens in {elapsed:.1f}s")

    # ── Step 7: Decode → waveform ──
    print(f"\nStep 6: conditional_decoder")
    speech_tok_arr = np.array([speech_tokens], dtype=np.int64)
    cd_out = sess_cd.run(None, {
        "speech_tokens":     speech_tok_arr,
        "speaker_embeddings": speaker_emb,
    })
    wav = cd_out[0].squeeze().astype(np.float32)

    # Normalize (mirrors browser worker)
    peak = np.abs(wav).max()
    if peak < 0.01:
        print(f"  warning: very low amplitude (peak={peak:.4f}), auto-normalizing")
        wav = wav * (0.9 / peak)
    wav = np.clip(wav, -1.0, 1.0)

    out_path = str(OUT_DIR / "browser_sim_output.wav")
    sf.write(out_path, wav, SAMPLE_RATE)
    print(f"\nSaved: {out_path}  ({len(wav)/SAMPLE_RATE:.2f}s, peak={np.abs(wav).max():.4f})")

    # ── Transcribe ──
    if GROQ_KEY:
        print("\nTranscribing with Groq Whisper...")
        with open(out_path, "rb") as f:
            r = requests.post(
                "https://api.groq.com/openai/v1/audio/transcriptions",
                headers={"Authorization": f"Bearer {GROQ_KEY}"},
                files={"file": (os.path.basename(out_path), f, "audio/wav")},
                data={"model": "whisper-large-v3", "language": "fi", "response_format": "text"},
            )
        if r.ok:
            print(f"\nTranscript:  '{r.text.strip()}'")
            print(f"Target text: '{TEXT}'")
        else:
            print(f"  Groq error: {r.status_code} {r.text}")


if __name__ == "__main__":
    main()

	"""
	browser_pipeline_sim.py

	Python simulation of the browser worker.ts logic.
	Mirrors the EXACT token assembly, CFG loop, sampling, and decoding
	that the browser worker does — making it trivial to diff against worker.ts.

	If this produces correct audio, worker.ts with the same logic will too.
	If there's a bug in the browser, you can find it by diffing this script against worker.ts.

	Usage:
	cd /workspaces/work
	conda run -n chatterbox-onnx python browser_pipeline_sim.py [text] [ref_audio]

	Outputs:
	_cmp/browser_sim_output.wav
	"""

	import os, sys, time, json, struct, math
	import numpy as np
	import onnxruntime as ort
	import soundfile as sf
	import librosa
	import requests
	from pathlib import Path
	from huggingface_hub import hf_hub_download

	sys.path.insert(0, "Chatterbox-Finnish")

	# ── Config — mirrors browser worker constants ─────────────────────────────────
	TEXT = sys.argv[1] if len(sys.argv) > 1 else \
	"Tervetuloa kokeilemaan hienoviritettyä suomenkielistä Chatterbox-puhesynteesiä."
	REF_AUDIO = sys.argv[2] if len(sys.argv) > 2 else \
	"Chatterbox-Finnish/samples/reference_finnish.wav"

	HF_BASE = "onnx-community/chatterbox-multilingual-ONNX"
	HF_FI = "RASMUS/Chatterbox-Finnish-ONNX"
	CACHE_DIR = "_onnx_cache"
	OUT_DIR = Path("_cmp"); OUT_DIR.mkdir(exist_ok=True)

	# Worker constants (from worker.ts)
	SOT = 255 # [START] token
	EOT = 0 # [STOP] token
	START_SPEECH = 6561 # BOS speech token
	STOP_SPEECH = 6562 # EOS speech token
	CFG_WEIGHT = 0.5
	REP_PENALTY = 1.2
	TEMPERATURE = 0.8
	EXAGGERATION = 0.6
	MIN_SPEECH_TOKENS = 40
	MAX_STEPS = 1000
	SAMPLE_RATE = 24000

	GROQ_KEY = os.environ.get("GROQ_API_KEY", "")

	# ── puncNorm — mirrors worker.ts puncNorm() ───────────────────────────────────
	def punc_norm(text: str) -> str:
	"""Mirrors puncNorm() in worker.ts"""
	import re
	t = text.strip()
	t = re.sub(r'\s+', ' ', t) # normalize whitespace
	t = t[0].upper() + t[1:] if t else t # capitalize first letter
	t = t.replace(' .', '.').replace(' ,', ',') # remove space before punct
	t = t.replace(' ?', '?').replace(' !', '!')
	if t and t[-1] not in '.!?…': # ensure ending punctuation
	t += '.'
	return t


	# ── EnTokenizer (same as Chatterbox-Finnish/src/chatterbox_/models/tokenizers/tokenizer.py)
	def load_tokenizer(tokenizer_path: str):
	from src.chatterbox_.models.tokenizers.tokenizer import EnTokenizer
	return EnTokenizer(tokenizer_path)


	# ── puncNorm + tokenize → mirrors worker.ts tokenize() ───────────────────────
	def tokenize(tok, text: str) -> list[int]:
	"""
	Mirrors browser worker tokenize():
	1. puncNorm
	2. replace spaces with [SPACE] token
	3. encode grapheme by grapheme
	4. wrap with [SOT, ..., EOT]
	"""
	normed = punc_norm(text)
	ids = tok.encode(normed)
	return [SOT] + ids + [EOT]


	# ── Download helpers ───────────────────────────────────────────────────────────
	def dl(repo_id, filename):
	return hf_hub_download(repo_id=repo_id, filename=filename,
	local_dir=CACHE_DIR, local_dir_use_symlinks=False)


	# ── KV cache helpers ───────────────────────────────────────────────────────────
	def empty_kv(n_layers=30, kv_dtype=np.float32):
	return [(np.zeros((1, 16, 0, 64), dtype=kv_dtype),
	np.zeros((1, 16, 0, 64), dtype=kv_dtype))
	for _ in range(n_layers)]


	def make_kv_feeds(kv_cache):
	feeds = {}
	for i, (k, v) in enumerate(kv_cache):
	feeds[f"past_key_values.{i}.key"] = k
	feeds[f"past_key_values.{i}.value"] = v
	return feeds


	# ── Sampling helpers ───────────────────────────────────────────────────────────
	def apply_rep_penalty(logits, generated_set, penalty):
	logits = logits.copy()
	for tok in generated_set:
	logits[tok] = logits[tok] / penalty if logits[tok] > 0 else logits[tok] * penalty
	return logits


	def apply_min_p(logits, p=0.05):
	"""Mirrors applyMinP() in worker.ts"""
	logits = logits.copy()
	probs = np.exp(logits - logits.max())
	probs /= probs.sum()
	threshold = probs.max() * p
	logits[probs < threshold] = -1e9
	return logits


	def sample_with_temperature(logits, temperature):
	"""Mirrors sampleWithTemperature() in worker.ts"""
	logits = (logits / temperature).astype(np.float64)
	logits -= logits.max()
	probs = np.exp(logits)
	probs /= probs.sum()
	return int(np.random.choice(len(probs), p=probs))


	# ── Main pipeline ─────────────────────────────────────────────────────────────
	def main():
	providers = ["CUDAExecutionProvider", "CPUExecutionProvider"]
	opts = ort.SessionOptions()
	opts.graph_optimization_level = ort.GraphOptimizationLevel.ORT_ENABLE_ALL

	# ── Load sessions ──
	print("Loading ONNX sessions...")
	sess_se = ort.InferenceSession(dl(HF_BASE, "onnx/speech_encoder.onnx"), opts, providers)
	sess_et = ort.InferenceSession(dl(HF_BASE, "onnx/embed_tokens.onnx"), opts, providers)
	sess_lm = ort.InferenceSession(dl(HF_FI, "onnx/language_model.onnx"), opts, providers)
	sess_cd = ort.InferenceSession(dl(HF_BASE, "onnx/conditional_decoder.onnx"), opts, providers)

	# ── Precomputed Finnish conditioning ──
	cond_emb_raw = open(dl(HF_FI, "onnx/finnish_cond_emb.bin"), "rb").read()
	cond_emb = np.frombuffer(cond_emb_raw, dtype=np.float32).reshape(1, 34, 1024)
	print(f" cond_emb: {cond_emb.shape}")

	# ── KV dtype from model ──
	kv_input_name = "past_key_values.0.key"
	kv_dtype_str = next(
	(inp.type for inp in sess_lm.get_inputs() if inp.name == kv_input_name), "tensor(float)"
	)
	kv_dtype = np.float16 if "float16" in kv_dtype_str else np.float32
	print(f" KV cache dtype: {kv_dtype}")

	# ── Step 1: Reference audio → speaker embeddings ──
	print(f"\nStep 1: speech_encoder ({REF_AUDIO})")
	ref_audio, ref_sr = librosa.load(REF_AUDIO, sr=None)
	ref_16k = librosa.resample(ref_audio, orig_sr=ref_sr, target_sr=16000).astype(np.float32)
	ref_input = ref_16k[np.newaxis, :] # [1, T]

	se_outs = sess_se.run(None, {"audio": ref_input})
	speaker_emb = se_outs[0] # [1, 256]
	prompt_tokens = se_outs[1] # [1, N]
	print(f" speaker_emb: {speaker_emb.shape}")
	print(f" prompt_tokens: {prompt_tokens.shape}")
	# Note: se_outs[2] is cond_emb from base model — we ignore it, use precomputed Finnish version

	# ── Step 2: Tokenize text (mirrors worker.ts tokenize()) ──
	print(f"\nStep 2: tokenize '{TEXT[:60]}...'")
	tok = load_tokenizer("Chatterbox-Finnish/pretrained_models/tokenizer.json")
	token_ids = tokenize(tok, TEXT)
	print(f" token_ids ({len(token_ids)}): {token_ids[:6]}...{token_ids[-3:]}")
	text_ids_np = np.array([token_ids], dtype=np.int64) # [1, T]

	# ── Step 3: Embed text tokens ──
	print(f"\nStep 3: embed_tokens")
	text_embeds = sess_et.run(None, {"input_ids": text_ids_np})[0] # [1, T, 1024]
	print(f" text_embeds: {text_embeds.shape}")

	# Embed BOS speech token
	bos_ids = np.array([[START_SPEECH]], dtype=np.int64)
	bos_emb = sess_et.run(None, {"input_ids": bos_ids})[0] # [1, 1, 1024]

	# ── Step 4: Build prefill ──
	# Matches PyTorch: inputs_embeds = cat([cond_emb, text_emb, bos_emb])
	prefill_cond = np.concatenate([cond_emb, text_embeds, bos_emb], axis=1)
	zeros_text = np.zeros_like(text_embeds)
	prefill_uncond = np.concatenate([cond_emb, zeros_text, bos_emb], axis=1)

	mask_cond = np.ones((1, prefill_cond.shape[1]), dtype=np.int64)
	mask_uncond = np.ones((1, prefill_uncond.shape[1]), dtype=np.int64)

	kv_empty_layer = np.zeros((1, 16, 0, 64), dtype=kv_dtype)
	kv_cond = [(kv_empty_layer.copy(), kv_empty_layer.copy()) for _ in range(30)]
	kv_uncond = [(kv_empty_layer.copy(), kv_empty_layer.copy()) for _ in range(30)]

	def lm_step(embeds, mask, kv):
	feeds = {"inputs_embeds": embeds, "attention_mask": mask}
	feeds.update(make_kv_feeds(kv))
	outs = sess_lm.run(None, feeds)
	logits = outs[0] # [1, seq, vocab]
	new_kv = [(outs[1 + i2], outs[1 + i2 + 1]) for i in range(30)]
	return logits, new_kv

	# ── Step 5: Prefill both streams ──
	print(f"\nStep 4: prefill ({prefill_cond.shape[1]} tokens)")
	t0 = time.time()
	logits_c, kv_cond = lm_step(prefill_cond, mask_cond, kv_cond)
	logits_uc, kv_uncond = lm_step(prefill_uncond, mask_uncond, kv_uncond)
	print(f" prefill done ({time.time()-t0:.1f}s)")

	# ── Step 6: Autoregressive generation ──
	print(f"\nStep 5: generate (max {MAX_STEPS} steps)")
	generated = [START_SPEECH]
	speech_tokens = []
	t0 = time.time()

	for step in range(MAX_STEPS):
	last_ids = np.array([[generated[-1]]], dtype=np.int64)
	last_emb = sess_et.run(None, {"input_ids": last_ids})[0] # [1, 1, 1024]

	seq_len_c = kv_cond[0][0].shape[2] + 1
	seq_len_uc = kv_uncond[0][0].shape[2] + 1
	mask_c = np.ones((1, seq_len_c), dtype=np.int64)
	mask_uc = np.ones((1, seq_len_uc), dtype=np.int64)

	logits_c, kv_cond = lm_step(last_emb, mask_c, kv_cond)
	logits_uc, kv_uncond = lm_step(last_emb, mask_uc, kv_uncond)

	# CFG: final = cond + cfg_weight * (cond - uncond)
	lc = logits_c[0, -1].astype(np.float32)
	luc = logits_uc[0, -1].astype(np.float32)
	final_logits = lc + CFG_WEIGHT * (lc - luc)

	# Apply rep penalty + min_p + temperature sample
	final_logits = apply_rep_penalty(final_logits, set(generated), REP_PENALTY)
	final_logits = apply_min_p(final_logits, p=0.05)
	token = sample_with_temperature(final_logits, TEMPERATURE)

	if token == STOP_SPEECH and len(speech_tokens) >= MIN_SPEECH_TOKENS:
	print(f" EOS at step {step} ({len(speech_tokens)} speech tokens)")
	break

	generated.append(token)
	if token < START_SPEECH:
	speech_tokens.append(token)

	if (step + 1) % 100 == 0:
	elapsed = time.time() - t0
	rate = (step + 1) / elapsed
	print(f" step {step+1:4d}: {len(speech_tokens):3d} speech tokens ({rate:.1f} tok/s)")

	elapsed = time.time() - t0
	print(f" generation done: {len(speech_tokens)} speech tokens in {elapsed:.1f}s")

	# ── Step 7: Decode → waveform ──
	print(f"\nStep 6: conditional_decoder")
	speech_tok_arr = np.array([speech_tokens], dtype=np.int64)
	cd_out = sess_cd.run(None, {
	"speech_tokens": speech_tok_arr,
	"speaker_embeddings": speaker_emb,
	})
	wav = cd_out[0].squeeze().astype(np.float32)

	# Normalize (mirrors browser worker)
	peak = np.abs(wav).max()
	if peak < 0.01:
	print(f" warning: very low amplitude (peak={peak:.4f}), auto-normalizing")
	wav = wav * (0.9 / peak)
	wav = np.clip(wav, -1.0, 1.0)

	out_path = str(OUT_DIR / "browser_sim_output.wav")
	sf.write(out_path, wav, SAMPLE_RATE)
	print(f"\nSaved: {out_path} ({len(wav)/SAMPLE_RATE:.2f}s, peak={np.abs(wav).max():.4f})")

	# ── Transcribe ──
	if GROQ_KEY:
	print("\nTranscribing with Groq Whisper...")
	with open(out_path, "rb") as f:
	r = requests.post(
	"https://api.groq.com/openai/v1/audio/transcriptions",
	headers={"Authorization": f"Bearer {GROQ_KEY}"},
	files={"file": (os.path.basename(out_path), f, "audio/wav")},
	data={"model": "whisper-large-v3", "language": "fi", "response_format": "text"},
	)
	if r.ok:
	print(f"\nTranscript: '{r.text.strip()}'")
	print(f"Target text: '{TEXT}'")
	else:
	print(f" Groq error: {r.status_code} {r.text}")


	if __name__ == "__main__":
	main()