echo-tts-preview / inference.py
vicentemovil's picture
Fix k_norm application in KV cache building
aa941cf
Raw
History Blame Contribute Delete
19.3 kB
from dataclasses import dataclass
from typing import Callable, List, Tuple
import torch
import safetensors.torch as st
from huggingface_hub import hf_hub_download
from model import EchoDiT
from autoencoder import build_ae, DAC
import torchaudio
from torchcodec.decoders import AudioDecoder
# from samplers import Sampler
SampleFn = Callable[
[EchoDiT, torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor, int],
torch.Tensor
]
@dataclass
class SpeakerCache:
"""Cached speaker encoder output and per-layer KV projections for reuse across text prompts."""
speaker_state: torch.Tensor # (1, 640, 1280) - speaker encoder output
speaker_mask: torch.Tensor # (1, 640) - patched mask
speaker_kvs: List[Tuple[torch.Tensor, torch.Tensor]] # 24 layers × (k, v)
### Loading
def load_model_from_hf(repo_id: str = 'jordand/echo-tts-base', device: str = 'cuda', dtype: torch.dtype | None = torch.bfloat16, compile: bool = False, token: str | None = None) -> EchoDiT:
with torch.device('meta'):
model = EchoDiT(
latent_size=80, model_size=2048, num_layers=24, num_heads=16,
intermediate_size=5888, norm_eps=1e-5, max_seq_len=640,
text_vocab_size=256, text_model_size=1280, text_num_layers=14,
text_num_heads=10, text_intermediate_size=3328, text_max_seq_len=768,
speaker_patch_size=4, speaker_model_size=1280, speaker_num_layers=14,
speaker_num_heads=10, speaker_intermediate_size=3328,
speaker_max_patched_seq_len=640, timestep_embed_size=512, adaln_rank=256,
)
w_path = hf_hub_download(repo_id, 'pytorch_model.safetensors', token=token)
# Load to CPU first
state = st.load_file(w_path, device='cpu')
# Convert dtype on CPU if needed
if dtype is not None:
state = {k: v.to(dtype=dtype) for k, v in state.items()}
# Now move to device
state = {k: v.to(device=device) for k, v in state.items()}
model.load_state_dict(state, strict=False, assign=True)
model = model.eval()
if compile:
model = torch.compile(model)
model.get_kv_cache = torch.compile(model.get_kv_cache)
return model
def load_fish_ae_from_hf(repo_id: str = 'jordand/fish-s1-dac-min', device: str = 'cuda', dtype: torch.dtype | None = torch.float32, compile: bool = False, token: str | None = None) -> DAC:
# have not tested lower precisions with fish AE yet
with torch.device('meta'):
fish_ae = build_ae()
w_path = hf_hub_download(repo_id, 'pytorch_model.safetensors', token=token)
# Load to CPU first (required for ZeroGPU compatibility)
state = st.load_file(w_path, device='cpu')
# Convert dtype if needed
if dtype is not None:
state = {k: v.to(dtype=dtype) for k, v in state.items()}
# Move to device
state = {k: v.to(device=device) for k, v in state.items()}
fish_ae.load_state_dict(state, strict=False, assign=True)
fish_ae = fish_ae.eval().to(device)
if compile:
fish_ae.encoder = torch.compile(fish_ae.encoder)
fish_ae.decoder = torch.compile(fish_ae.decoder)
return fish_ae
@dataclass
class PCAState:
pca_components: torch.Tensor
pca_mean: torch.Tensor
latent_scale: float
def load_pca_state_from_hf(repo_id: str = 'jordand/echo-tts-base', device: str = 'cuda', filename: str = 'pca_state.safetensors', token: str | None = None) -> PCAState:
p_path = hf_hub_download(repo_id, filename, token=token)
# Load to CPU first (required for ZeroGPU compatibility)
t = st.load_file(p_path, device='cpu')
# Move tensors to device
t = {k: v.to(device) for k, v in t.items()}
return PCAState(
pca_components=t["pca_components"],
pca_mean=t["pca_mean"],
latent_scale=float(t["latent_scale"].item()),
)
### default load audio
def load_audio(path: str) -> torch.Tensor:
decoder = AudioDecoder(path)
sr = decoder.metadata.sample_rate
audio = decoder.get_samples_played_in_range(0, 120)
audio = audio.data.mean(dim=0).unsqueeze(0)
audio = torchaudio.functional.resample(audio, sr, 44_100)
audio = audio / torch.maximum(audio.abs().max(), torch.tensor(1.))
# TODO is this better than clipping? should we target a specific energy level?
return audio
### Text helpers
def tokenizer_encode(text: str, append_bos: bool = True, normalize: bool = True) -> torch.Tensor:
if normalize:
text = text.replace('…', '...')
text = text.replace('“', '"')
text = text.replace('”', '"')
text = text.replace('’', "'")
text = text.replace('\n', " ")
text = text.replace(':', ',')
text = text.replace(';', ',')
b = list(text.encode('utf-8'))
if append_bos:
b.insert(0, 0)
return torch.tensor(b)
def get_text_input_ids_and_mask(text_arr: List[str], max_length: int | None, device: str | None = None) -> tuple[torch.Tensor, torch.Tensor]:
batch_size = len(text_arr)
if max_length is None:
max_length = max(len(tokenizer_encode(text)) for text in text_arr) # obviously bad...
tokens = torch.zeros((batch_size, max_length), dtype=torch.int32)
mask = torch.zeros((batch_size, max_length), dtype=torch.bool)
for i, text in enumerate(text_arr):
encoded = tokenizer_encode(text)
length = min(len(encoded), max_length)
tokens[i, :length] = encoded[:length]
mask[i, :length] = 1
if device is not None:
tokens = tokens.to(device)
mask = mask.to(device)
return tokens, mask
### Autoencoder Inference
@torch.inference_mode()
def ae_encode(fish_ae: DAC, pca_state: PCAState, audio: torch.Tensor) -> torch.Tensor:
assert audio.ndim == 3 and audio.shape[1] == 1 # (b, 1, length)
z_q = fish_ae.encode_zq(audio).float()
z_q = (z_q.transpose(1, 2) - pca_state.pca_mean) @ pca_state.pca_components.T
z_q = z_q * pca_state.latent_scale
return z_q
@torch.inference_mode()
def ae_decode(fish_ae: DAC, pca_state: PCAState, z_q: torch.Tensor) -> torch.Tensor:
z_q = (z_q / pca_state.latent_scale) @ pca_state.pca_components + pca_state.pca_mean
return fish_ae.decode_zq(z_q.transpose(1, 2).to(fish_ae.dtype)).float()
@torch.inference_mode()
def ae_reconstruct(fish_ae: DAC, pca_state: PCAState, audio: torch.Tensor) -> torch.Tensor:
# (audio is (b, 1, length))
z_q = ae_encode(fish_ae, pca_state, audio.to(fish_ae.dtype))
return ae_decode(fish_ae, pca_state, z_q)
@torch.inference_mode()
def get_speaker_latent_and_mask(
fish_ae: DAC,
pca_state: PCAState,
audio: torch.Tensor, # (1, length)
max_speaker_latent_len: int = 2560, # pretrained max length
audio_chunk_size: int = 640 * 2048 # (~30 seconds, 1/4 max speaker condition size)
) -> tuple[torch.Tensor, torch.Tensor]:
# gets speaker latent and mask from audio, computes in chunks and concatenates (similar to pretraining setup)
AE_DOWNSAMPLE_FACTOR = 2048
max_audio_len = max_speaker_latent_len * AE_DOWNSAMPLE_FACTOR
assert audio.ndim == 2 and audio.shape[0] == 1 # (1, length)
audio = audio[:, :max_audio_len]
audio_len = audio.shape[1]
latent_arr = []
for i in range(0, audio_len, audio_chunk_size):
audio_chunk = audio[:, i:i + audio_chunk_size]
if audio_chunk.shape[1] < audio_chunk_size:
audio_chunk = torch.nn.functional.pad(audio_chunk, (0, audio_chunk_size - audio_chunk.shape[1]))
latent_chunk = ae_encode(fish_ae, pca_state, audio_chunk.unsqueeze(0))
latent_arr.append(latent_chunk)
speaker_latent = torch.cat(latent_arr, dim=1)
actual_latent_len = audio_len // AE_DOWNSAMPLE_FACTOR
speaker_mask = (torch.arange(speaker_latent.shape[1], device=speaker_latent.device) < actual_latent_len).unsqueeze(0)
if speaker_latent.shape[1] < max_speaker_latent_len:
speaker_latent = torch.nn.functional.pad(speaker_latent, (0, 0, 0, max_speaker_latent_len - speaker_latent.shape[1]))
speaker_mask = torch.nn.functional.pad(speaker_mask, (0, max_speaker_latent_len - speaker_mask.shape[1]))
return speaker_latent, speaker_mask
### Speaker cache functions for batch generation
@torch.inference_mode()
def build_speaker_cache(
model: EchoDiT,
fish_ae: DAC,
pca_state: PCAState,
speaker_audio: torch.Tensor,
pad_to_max_speaker_latent_len: int = 2560
) -> SpeakerCache:
"""Build speaker cache: encode speaker audio once and project into all layer KVs."""
speaker_latent, speaker_mask = get_speaker_latent_and_mask(
fish_ae, pca_state,
speaker_audio.to(fish_ae.dtype),
max_speaker_latent_len=pad_to_max_speaker_latent_len
)
speaker_state = model.speaker_encoder(speaker_latent.to(model.dtype))
speaker_state = model.speaker_norm(speaker_state)
speaker_mask_patched = speaker_mask[..., ::model.speaker_patch_size]
# Project speaker_state into per-layer K/V and apply k_norm
speaker_kvs = []
for block in model.blocks:
attn = block.attention
bs = speaker_state.shape[0]
xk = attn.wk_speaker(speaker_state).reshape(bs, speaker_state.shape[1], attn.num_heads, -1)
xv = attn.wv_speaker(speaker_state).reshape(bs, speaker_state.shape[1], attn.num_heads, -1)
xk = attn.k_norm(xk) # Apply k_norm before caching
speaker_kvs.append((xk, xv))
return SpeakerCache(
speaker_state=speaker_state,
speaker_mask=speaker_mask_patched,
speaker_kvs=speaker_kvs
)
@torch.inference_mode()
def get_zero_speaker_cache(model: EchoDiT, device: str = 'cuda') -> SpeakerCache:
"""Compute speaker cache for zero speaker (all-zeros latent) - computed once at startup."""
speaker_latent_zero = torch.zeros((1, 2560, 80), device=device, dtype=model.dtype)
speaker_state_zero = model.speaker_encoder(speaker_latent_zero)
speaker_state_zero = model.speaker_norm(speaker_state_zero)
speaker_mask_zero = torch.zeros((1, 640), device=device, dtype=torch.bool)
speaker_kvs = []
for block in model.blocks:
attn = block.attention
bs = 1
xk = attn.wk_speaker(speaker_state_zero).reshape(bs, speaker_state_zero.shape[1], attn.num_heads, -1)
xv = attn.wv_speaker(speaker_state_zero).reshape(bs, speaker_state_zero.shape[1], attn.num_heads, -1)
xk = attn.k_norm(xk) # Apply k_norm before caching
speaker_kvs.append((xk, xv))
return SpeakerCache(
speaker_state=speaker_state_zero,
speaker_mask=speaker_mask_zero,
speaker_kvs=speaker_kvs
)
### Full sample pipeline
def find_flattening_point(data, target_value=0.0, window_size=20, std_threshold=0.05):
padded_data = torch.cat([data, torch.zeros(window_size, *data.shape[1:], device=data.device, dtype=data.dtype)])
for i in range(len(padded_data) - window_size):
window = padded_data[i:i + window_size]
if window.std() < std_threshold and abs(window.mean() - target_value) < 0.1:
return i
return len(data)
@torch.inference_mode()
def sample_pipeline(
model: EchoDiT,
fish_ae: DAC,
pca_state: PCAState,
sample_fn: SampleFn,
text_prompt: str,
speaker_audio: torch.Tensor | None,
rng_seed: int,
pad_to_max_speaker_latent_len: int | None = 2560,
pad_to_max_text_seq_len: int | None = 768,
) -> torch.Tensor | tuple[torch.Tensor, torch.Tensor]:
MAX_SPEAKER_LATENT_LEN = 2560
MAX_TEXT_SEQ_LEN = 768
device, dtype = model.device, model.dtype
text_input_ids, text_mask = get_text_input_ids_and_mask([text_prompt], min(pad_to_max_text_seq_len or MAX_TEXT_SEQ_LEN, MAX_TEXT_SEQ_LEN), device=device)
# print('initial text input ids length: ', text_input_ids.shape[1])
# torch.cuda.synchronize()
# import time
# t0 = time.time()
if speaker_audio is None:
# No speaker prompt - use zero speaker latent and mask
speaker_latent = torch.zeros((1, pad_to_max_speaker_latent_len if pad_to_max_speaker_latent_len else MAX_SPEAKER_LATENT_LEN, 80), device=device, dtype=dtype)
speaker_mask = torch.zeros((1, pad_to_max_speaker_latent_len if pad_to_max_speaker_latent_len else MAX_SPEAKER_LATENT_LEN), device=device, dtype=torch.bool)
# print("Using zero speaker latent and mask (no speaker prompt)")
else:
speaker_latent, speaker_mask = get_speaker_latent_and_mask(
fish_ae,
pca_state,
speaker_audio.to(fish_ae.dtype),
max_speaker_latent_len=pad_to_max_speaker_latent_len if pad_to_max_speaker_latent_len else MAX_SPEAKER_LATENT_LEN
)
speaker_latent = speaker_latent.to(device)
speaker_mask = speaker_mask.to(device)
# print('speaker latent shape: ', speaker_latent.shape)
# print('speaker mask shape: ', speaker_mask.shape)
# torch.cuda.synchronize()
# t1 = time.time()
# print(f"Time taken encode: {t1 - t0} seconds")
latent_out = sample_fn(model, speaker_latent, speaker_mask, text_input_ids, text_mask, rng_seed)
# torch.cuda.synchronize()
# t2 = time.time()
# print(f"Time taken sample: {t2 - t1} seconds")
audio_out = ae_decode(fish_ae, pca_state, latent_out)
# torch.cuda.synchronize()
# t3 = time.time()
# print(f"Time taken decode: {t3 - t2} seconds")
flattening_point = find_flattening_point(latent_out[0])
audio_out = audio_out[..., :flattening_point * 2048]
# print(f"\nTime taken total: {t3 - t0} seconds")
# peak_mem = torch.cuda.max_memory_allocated()
# print(f"Peak memory: {peak_mem / 1024**2:.2f} MB")
# print(torch.cuda.memory_summary(abbreviated=True))
return audio_out
@torch.inference_mode()
def build_kv_cache_from_speaker_cache(
model: EchoDiT,
speaker_cache: SpeakerCache,
zero_speaker_cache: SpeakerCache,
text_input_ids: torch.Tensor,
text_mask: torch.Tensor,
guidance_mode: str = "independent"
) -> tuple[List[Tuple[torch.Tensor, torch.Tensor]], torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
"""Build full KV cache for CFG from cached speaker KVs + fresh text encoding.
Returns: (kv_cache, combined_text_mask, combined_speaker_mask, text_mask, text_mask_uncond)
The last two are for alternating mode which needs per-step switching.
"""
device = model.device
# Text encoding - the only variable cost per text
text_state = model.text_encoder(text_input_ids, text_mask)
text_state = model.text_norm(text_state)
# Unconditional text (BOS only)
text_input_ids_uncond = torch.zeros_like(text_input_ids)
text_mask_uncond = torch.zeros_like(text_mask)
text_mask_uncond[:, 0] = True
text_state_uncond = model.text_encoder(text_input_ids_uncond, text_mask_uncond)
text_state_uncond = model.text_norm(text_state_uncond)
# Assemble CFG batch based on guidance mode
kv_cache = []
for i, block in enumerate(model.blocks):
attn = block.attention
bs = text_state.shape[0]
def _text_kv(ts):
xk = attn.wk_text(ts).reshape(bs, ts.shape[1], attn.num_heads, -1)
xv = attn.wv_text(ts).reshape(bs, ts.shape[1], attn.num_heads, -1)
xk = attn.k_norm(xk) # Apply k_norm to text KVs
return xk, xv
xk_txt_cond, xv_txt_cond = _text_kv(text_state)
xk_txt_uncond, xv_txt_uncond = _text_kv(text_state_uncond)
# Speaker KVs are already k_norm'd in the cache
xk_sp_cond, xv_sp_cond = speaker_cache.speaker_kvs[i]
xk_sp_zero, xv_sp_zero = zero_speaker_cache.speaker_kvs[i]
if guidance_mode in ["independent", "apg-independent"]:
# 3× batch: [cond, text_uncond, speaker_uncond]
xk = torch.cat([
torch.cat([xk_txt_cond, xk_sp_cond], dim=1), # batch 0: cond
torch.cat([xk_txt_uncond, xk_sp_cond], dim=1), # batch 1: text-uncond
torch.cat([xk_txt_cond, xk_sp_zero], dim=1), # batch 2: speaker-uncond
], dim=0)
xv = torch.cat([
torch.cat([xv_txt_cond, xv_sp_cond], dim=1),
torch.cat([xv_txt_uncond, xv_sp_cond], dim=1),
torch.cat([xv_txt_cond, xv_sp_zero], dim=1),
], dim=0)
elif guidance_mode == "alternating":
# 2× batch: [cond, cond] - note: using cond text in both, will switch masks per step
# This follows the original implementation which has known limitations (see samplers.py line 337):
# Uses text_cond with alternating masks rather than true text_uncond encoding.
# Works because text encoder is bidirectional (not causal), but is "technically incorrect"
# per original author. In practice: "might not make much of a difference."
# No presets use alternating mode. Could be improved by building separate KV caches per step,
# but that would defeat the speaker cache optimization.
xk = torch.cat([
torch.cat([xk_txt_cond, xk_sp_cond], dim=1), # batch 0: cond
torch.cat([xk_txt_cond, xk_sp_zero], dim=1), # batch 1: text_cond + speaker_zero (masks will alternate)
], dim=0)
xv = torch.cat([
torch.cat([xv_txt_cond, xv_sp_cond], dim=1),
torch.cat([xv_txt_cond, xv_sp_zero], dim=1),
], dim=0)
else: # joint-unconditional
# 2× batch: [cond, joint_uncond]
xk = torch.cat([
torch.cat([xk_txt_cond, xk_sp_cond], dim=1), # batch 0: cond
torch.cat([xk_txt_uncond, xk_sp_zero], dim=1), # batch 1: joint uncond
], dim=0)
xv = torch.cat([
torch.cat([xv_txt_cond, xv_sp_cond], dim=1),
torch.cat([xv_txt_uncond, xv_sp_zero], dim=1),
], dim=0)
# k_norm already applied to both text and speaker KVs before concatenation
kv_cache.append((xk, xv))
# Combined masks based on mode
# When using prebuilt kv_cache, we need to return a single combined cross-attention mask
# (text + speaker concatenated), not separate masks
if guidance_mode in ["independent", "apg-independent"]:
text_masks = torch.cat([text_mask, text_mask_uncond, text_mask], dim=0)
speaker_masks = torch.cat([
speaker_cache.speaker_mask,
speaker_cache.speaker_mask,
zero_speaker_cache.speaker_mask
], dim=0)
else: # alternating or joint
text_masks = torch.cat([text_mask, text_mask_uncond], dim=0)
speaker_masks = torch.cat([
speaker_cache.speaker_mask,
zero_speaker_cache.speaker_mask
], dim=0)
# Concatenate text and speaker masks into single cross-attention mask
combined_cross_mask = torch.cat([text_masks, speaker_masks], dim=1)
# Return empty mask for speaker_mask parameter (model will concat [self, cross, empty])
empty_mask = torch.zeros((combined_cross_mask.shape[0], 0), dtype=torch.bool, device=device)
return kv_cache, combined_cross_mask, empty_mask, text_mask, text_mask_uncond