Upload folder using huggingface_hub
Browse files- .gitattributes +1 -0
- .meta/create_model.py +81 -0
- .meta/env.json +34 -0
- .meta/model.log +0 -0
- README.md +70 -0
- config.json +107 -0
- model.safetensors +3 -0
- tokenizer.json +3 -0
.gitattributes
CHANGED
|
@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
|
|
| 33 |
*.zip filter=lfs diff=lfs merge=lfs -text
|
| 34 |
*.zst filter=lfs diff=lfs merge=lfs -text
|
| 35 |
*tfevents* filter=lfs diff=lfs merge=lfs -text
|
|
|
|
|
|
| 33 |
*.zip filter=lfs diff=lfs merge=lfs -text
|
| 34 |
*.zst filter=lfs diff=lfs merge=lfs -text
|
| 35 |
*tfevents* filter=lfs diff=lfs merge=lfs -text
|
| 36 |
+
tokenizer.json filter=lfs diff=lfs merge=lfs -text
|
.meta/create_model.py
ADDED
|
@@ -0,0 +1,81 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
def create_model(source_model_id: str, source_revision: str, save_folder: str) -> str:
|
| 2 |
+
import json
|
| 3 |
+
from pathlib import Path
|
| 4 |
+
|
| 5 |
+
import torch
|
| 6 |
+
from huggingface_hub import hf_hub_download
|
| 7 |
+
from safetensors.torch import save_file
|
| 8 |
+
|
| 9 |
+
folder = Path(save_folder)
|
| 10 |
+
folder.mkdir(parents=True, exist_ok=True)
|
| 11 |
+
raw = json.loads(Path(hf_hub_download(source_model_id, "config.json", revision=source_revision)).read_text())
|
| 12 |
+
|
| 13 |
+
hidden_size = 32
|
| 14 |
+
num_layers = 8
|
| 15 |
+
latent_dim = raw["codec"]["latent_dim"]
|
| 16 |
+
text = raw["lm"]
|
| 17 |
+
text.update(
|
| 18 |
+
hidden_size=hidden_size,
|
| 19 |
+
intermediate_size=2 * hidden_size,
|
| 20 |
+
num_hidden_layers=num_layers,
|
| 21 |
+
num_attention_heads=4,
|
| 22 |
+
num_key_value_heads=1,
|
| 23 |
+
head_dim=16,
|
| 24 |
+
linear_num_key_heads=2,
|
| 25 |
+
linear_num_value_heads=4,
|
| 26 |
+
linear_key_head_dim=16,
|
| 27 |
+
linear_value_head_dim=16,
|
| 28 |
+
max_position_embeddings=2048,
|
| 29 |
+
tie_word_embeddings=True,
|
| 30 |
+
)
|
| 31 |
+
# Keep the source 3-linear/1-full cadence for two complete cycles.
|
| 32 |
+
text["layer_types"] = list(text["layer_types"][:num_layers])
|
| 33 |
+
codec = raw["codec"]
|
| 34 |
+
codec.update(
|
| 35 |
+
base_channels=1,
|
| 36 |
+
encoder_depths=[1] * len(codec["encoder_depths"]),
|
| 37 |
+
decoder_depths=[1] * len(codec["decoder_depths"]),
|
| 38 |
+
)
|
| 39 |
+
raw["head"].update(hidden_size=hidden_size, num_layers=1)
|
| 40 |
+
raw["dtype"] = "bfloat16"
|
| 41 |
+
|
| 42 |
+
from haru.config import HaruConfig
|
| 43 |
+
from haru.models.haru import HaruModel
|
| 44 |
+
|
| 45 |
+
config = HaruConfig.from_dict(raw)
|
| 46 |
+
torch.manual_seed(42)
|
| 47 |
+
model = HaruModel(config)
|
| 48 |
+
with torch.no_grad():
|
| 49 |
+
for name, parameter in sorted(model.named_parameters()):
|
| 50 |
+
if name.endswith("bias"):
|
| 51 |
+
parameter.zero_()
|
| 52 |
+
elif "norm" in name and parameter.ndim == 1:
|
| 53 |
+
parameter.fill_(1)
|
| 54 |
+
elif name.endswith(("gamma", "weight_scale")):
|
| 55 |
+
parameter.fill_(1)
|
| 56 |
+
else:
|
| 57 |
+
# The codec allocates conv weights with torch.empty and never
|
| 58 |
+
# initializes them, so give every remaining tensor a deterministic
|
| 59 |
+
# value instead of leftover process memory.
|
| 60 |
+
parameter.normal_(0, 0.02)
|
| 61 |
+
|
| 62 |
+
state = {}
|
| 63 |
+
for name, tensor in sorted(model.state_dict().items()):
|
| 64 |
+
value = tensor.detach().cpu().contiguous()
|
| 65 |
+
state[name] = value.to(torch.bfloat16) if value.is_floating_point() else value
|
| 66 |
+
assert not any(name.startswith("lm_head") for name in state)
|
| 67 |
+
assert state["lm.embed_tokens.weight"].shape == (text["vocab_size"], hidden_size)
|
| 68 |
+
save_file(state, str(folder / "model.safetensors"), metadata={"format": "pt"})
|
| 69 |
+
(folder / "config.json").write_text(json.dumps(raw, indent=2) + "\n")
|
| 70 |
+
(folder / "tokenizer.json").write_bytes(
|
| 71 |
+
Path(hf_hub_download(source_model_id, "tokenizer.json", revision=source_revision)).read_bytes()
|
| 72 |
+
)
|
| 73 |
+
return str(model)
|
| 74 |
+
|
| 75 |
+
|
| 76 |
+
if __name__ == "__main__":
|
| 77 |
+
create_model(
|
| 78 |
+
source_model_id="higurashi-ai/Haru-4B",
|
| 79 |
+
source_revision="8d6e83bcd3701a76fc27a764965665da1df7da78",
|
| 80 |
+
save_folder="/tmp/haru",
|
| 81 |
+
)
|
.meta/env.json
ADDED
|
@@ -0,0 +1,34 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"source_model": {
|
| 3 |
+
"name": "higurashi-ai/Haru-4B",
|
| 4 |
+
"commit_id": "8d6e83bcd3701a76fc27a764965665da1df7da78"
|
| 5 |
+
},
|
| 6 |
+
"environments": {
|
| 7 |
+
"creation": {
|
| 8 |
+
"python": "3.12.13",
|
| 9 |
+
"packages": {
|
| 10 |
+
"huggingface_hub": "2.3.0",
|
| 11 |
+
"safetensors": "0.8.0",
|
| 12 |
+
"tokenizers": "0.23.3",
|
| 13 |
+
"torch": "2.13.0+cu126",
|
| 14 |
+
"transformers": "5.22.0.dev0"
|
| 15 |
+
},
|
| 16 |
+
"cuda": "12.6"
|
| 17 |
+
},
|
| 18 |
+
"haru": {
|
| 19 |
+
"python": "3.12.13",
|
| 20 |
+
"packages": {
|
| 21 |
+
"huggingface_hub": "2.3.0",
|
| 22 |
+
"numpy": "2.5.2",
|
| 23 |
+
"safetensors": "0.8.0",
|
| 24 |
+
"tokenizers": "0.23.3",
|
| 25 |
+
"torch": "2.13.0+cu126",
|
| 26 |
+
"transformers": "5.22.0.dev0"
|
| 27 |
+
},
|
| 28 |
+
"cuda": "12.6",
|
| 29 |
+
"hardware": {
|
| 30 |
+
"gpu_model": "NVIDIA H20"
|
| 31 |
+
}
|
| 32 |
+
}
|
| 33 |
+
}
|
| 34 |
+
}
|
.meta/model.log
ADDED
|
File without changes
|
README.md
ADDED
|
@@ -0,0 +1,70 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
library_name: haru
|
| 3 |
+
pipeline_tag: audio-to-audio
|
| 4 |
+
base_model:
|
| 5 |
+
- higurashi-ai/Haru-4B
|
| 6 |
+
---
|
| 7 |
+
|
| 8 |
+
This is a randomly initialized tiny model based on [higurashi-ai/Haru-4B](https://huggingface.co/higurashi-ai/Haru-4B) for testing and debugging.
|
| 9 |
+
|
| 10 |
+
| Feature | Source | Tiny |
|
| 11 |
+
|---|---|---|
|
| 12 |
+
| Checkpoint tensor elements | 5.27B BF16 | 8.36M BF16 |
|
| 13 |
+
| Safetensors weight files | 10.54 GB (model.safetensors) | 16.74 MB (model.safetensors) |
|
| 14 |
+
| Language model | Qwen3.5-4B, 32 layers | Qwen3.5, 8 layers |
|
| 15 |
+
| Attention cadence | 3 linear + 1 full | Same; two complete cycles |
|
| 16 |
+
| Hidden / MLP width | 2,560 / 9,216 | 32 / 64 |
|
| 17 |
+
| Q heads / KV heads / head dimension | 16 / 4 / 256 | 4 / 1 / 16 |
|
| 18 |
+
| Vocabulary / tied embeddings | 248,320 / true | 248,320 / true |
|
| 19 |
+
| Codec | VibeVoice, base 32, depths 3-8 | VibeVoice, base 1, depth 1 |
|
| 20 |
+
| Audio latents | 64 dims at 7.5 Hz | 64 dims at 7.5 Hz |
|
| 21 |
+
| MeanFlow head layers / steps | 4 / 1 | 1 / 1 |
|
| 22 |
+
|
| 23 |
+
| File path | Size |
|
| 24 |
+
|------|------|
|
| 25 |
+
| model.safetensors | 16.7MB |
|
| 26 |
+
|
| 27 |
+
### Example usage
|
| 28 |
+
|
| 29 |
+
#### Haru duplex session (✅ verified)
|
| 30 |
+
|
| 31 |
+
Install with `pip install 'git+https://github.com/higurashi-ai/Haru#subdirectory=haru'`. Runs six 133 ms frames with a synthetic reference voice; only the language model's linear-attention layers need CUDA, the codec and head also run on CPU.
|
| 32 |
+
|
| 33 |
+
```python
|
| 34 |
+
def run_haru(model_id: str) -> None:
|
| 35 |
+
import numpy as np
|
| 36 |
+
import torch
|
| 37 |
+
from haru import DuplexConfig, DuplexSession, HaruModel
|
| 38 |
+
|
| 39 |
+
device = "cuda" if torch.cuda.is_available() else "cpu"
|
| 40 |
+
model = HaruModel.from_pretrained(model_id, device=device)
|
| 41 |
+
seconds = np.arange(24_000) / 24_000
|
| 42 |
+
voice = (0.2 * np.sin(2 * np.pi * 150 * seconds)).astype(np.float32)
|
| 43 |
+
config = DuplexConfig(seed=7, speak_first=True, max_frames=6)
|
| 44 |
+
session = DuplexSession(model, "You are Haru. Greet the other person.", voice, config)
|
| 45 |
+
frames = [session.start()]
|
| 46 |
+
while not frames[-1].done and len(frames) < config.max_frames:
|
| 47 |
+
frames.append(session.step(np.zeros(3200, dtype=np.float32)))
|
| 48 |
+
assert len(frames) == 6
|
| 49 |
+
assert all(frame.audio.shape == (3200,) and frame.audio.dtype == np.float32 for frame in frames)
|
| 50 |
+
assert all(np.isfinite(frame.audio).all() and np.abs(frame.audio).max() <= 1.0 for frame in frames)
|
| 51 |
+
assert all(0 <= frame.token < 248320 for frame in frames)
|
| 52 |
+
assert frames[1].token not in (248076, 248071)
|
| 53 |
+
print("frames", len(frames), "text", repr("".join(frame.text for frame in frames)[:80]))
|
| 54 |
+
|
| 55 |
+
|
| 56 |
+
if __name__ == "__main__":
|
| 57 |
+
run_haru(model_id="tiny-random/haru")
|
| 58 |
+
```
|
| 59 |
+
|
| 60 |
+
### Creating this repo
|
| 61 |
+
|
| 62 |
+
- Code: [.meta/create_model.py](.meta/create_model.py)
|
| 63 |
+
- Model structure: [.meta/model.log](.meta/model.log)
|
| 64 |
+
- Environment: [.meta/env.json](.meta/env.json)
|
| 65 |
+
|
| 66 |
+
### Change log
|
| 67 |
+
|
| 68 |
+
- 2026-11-15: Tiny Haru full-duplex checkpoint with a shrunken Qwen3.5 backbone, codec and MeanFlow head.
|
| 69 |
+
|
| 70 |
+
*This repo is generated by GPT-6.1-Sol*
|
config.json
ADDED
|
@@ -0,0 +1,107 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"model_type": "haru",
|
| 3 |
+
"architectures": [
|
| 4 |
+
"HaruForConditionalGeneration"
|
| 5 |
+
],
|
| 6 |
+
"dtype": "bfloat16",
|
| 7 |
+
"sample_rate": 24000,
|
| 8 |
+
"frame_samples": 3200,
|
| 9 |
+
"frame_rate": 7.5,
|
| 10 |
+
"latent_dim": 64,
|
| 11 |
+
"pad_token_id": 248076,
|
| 12 |
+
"eop_token_id": 248071,
|
| 13 |
+
"lm": {
|
| 14 |
+
"model_type": "qwen3_5_text",
|
| 15 |
+
"vocab_size": 248320,
|
| 16 |
+
"hidden_size": 32,
|
| 17 |
+
"intermediate_size": 64,
|
| 18 |
+
"num_hidden_layers": 8,
|
| 19 |
+
"num_attention_heads": 4,
|
| 20 |
+
"num_key_value_heads": 1,
|
| 21 |
+
"head_dim": 16,
|
| 22 |
+
"hidden_act": "silu",
|
| 23 |
+
"max_position_embeddings": 2048,
|
| 24 |
+
"rms_norm_eps": 1e-06,
|
| 25 |
+
"attention_bias": false,
|
| 26 |
+
"attention_dropout": 0.0,
|
| 27 |
+
"attn_output_gate": true,
|
| 28 |
+
"full_attention_interval": 4,
|
| 29 |
+
"layer_types": [
|
| 30 |
+
"linear_attention",
|
| 31 |
+
"linear_attention",
|
| 32 |
+
"linear_attention",
|
| 33 |
+
"full_attention",
|
| 34 |
+
"linear_attention",
|
| 35 |
+
"linear_attention",
|
| 36 |
+
"linear_attention",
|
| 37 |
+
"full_attention"
|
| 38 |
+
],
|
| 39 |
+
"linear_conv_kernel_dim": 4,
|
| 40 |
+
"linear_key_head_dim": 16,
|
| 41 |
+
"linear_value_head_dim": 16,
|
| 42 |
+
"linear_num_key_heads": 2,
|
| 43 |
+
"linear_num_value_heads": 4,
|
| 44 |
+
"mamba_ssm_dtype": "float32",
|
| 45 |
+
"partial_rotary_factor": 0.25,
|
| 46 |
+
"rope_parameters": {
|
| 47 |
+
"mrope_interleaved": true,
|
| 48 |
+
"mrope_section": [
|
| 49 |
+
11,
|
| 50 |
+
11,
|
| 51 |
+
10
|
| 52 |
+
],
|
| 53 |
+
"partial_rotary_factor": 0.25,
|
| 54 |
+
"rope_theta": 10000000,
|
| 55 |
+
"rope_type": "default"
|
| 56 |
+
},
|
| 57 |
+
"tie_word_embeddings": true,
|
| 58 |
+
"initializer_range": 0.02,
|
| 59 |
+
"eos_token_id": 248044
|
| 60 |
+
},
|
| 61 |
+
"codec": {
|
| 62 |
+
"latent_dim": 64,
|
| 63 |
+
"base_channels": 1,
|
| 64 |
+
"ratios": [
|
| 65 |
+
8,
|
| 66 |
+
5,
|
| 67 |
+
5,
|
| 68 |
+
4,
|
| 69 |
+
2,
|
| 70 |
+
2
|
| 71 |
+
],
|
| 72 |
+
"encoder_depths": [
|
| 73 |
+
1,
|
| 74 |
+
1,
|
| 75 |
+
1,
|
| 76 |
+
1,
|
| 77 |
+
1,
|
| 78 |
+
1,
|
| 79 |
+
1
|
| 80 |
+
],
|
| 81 |
+
"decoder_depths": [
|
| 82 |
+
1,
|
| 83 |
+
1,
|
| 84 |
+
1,
|
| 85 |
+
1,
|
| 86 |
+
1,
|
| 87 |
+
1,
|
| 88 |
+
1
|
| 89 |
+
],
|
| 90 |
+
"norm_eps": 1e-05,
|
| 91 |
+
"latent_std": 0.5
|
| 92 |
+
},
|
| 93 |
+
"head": {
|
| 94 |
+
"hidden_size": 32,
|
| 95 |
+
"latent_dim": 64,
|
| 96 |
+
"num_layers": 1,
|
| 97 |
+
"ffn_ratio": 3.0,
|
| 98 |
+
"norm_eps": 1e-05,
|
| 99 |
+
"interval_embed": 7.0,
|
| 100 |
+
"num_steps": 1,
|
| 101 |
+
"guidance_scale": 1.5,
|
| 102 |
+
"guidance_range": [
|
| 103 |
+
1.0,
|
| 104 |
+
2.0
|
| 105 |
+
]
|
| 106 |
+
}
|
| 107 |
+
}
|
model.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:a677e4de10980183f931806dec5585a307e94aab29e94d755be86f6622b648ba
|
| 3 |
+
size 16741138
|
tokenizer.json
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:06b9509352d2af50381ab2247e083b80d32d5c0aba91c272ca9ff729b6a0e523
|
| 3 |
+
size 19989325
|