bhetherman/live-voice-chat
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

working ok

Browse Source
This commit is contained in:
Brian
2026-04-16 10:00:37 -04:00
parent 9debc56137
commit 129df7d1fa
24 changed files with 674 additions and 539 deletions
Download Patch File Download Diff File Expand all files Collapse all files
tests/README.md
+28 -8
View File
@@ -9,25 +9,45 @@ python -m pytest tests/unit -v
```
These exercise pure logic: config parsing, prompt derivation, LoRA spec
parsing, frame-length fitting, library round-robin selection. They do not
touch CUDA, Wan2.2, MuseTalk, or ffmpeg. Safe to run on Windows, outside
Docker, without any models installed.
parsing, frame-length fitting, library round-robin selection, the
pipeline's video branch, and ffmpeg mux argument shaping. They do not
touch CUDA, Wan2.2, MuseTalk, or a real ffmpeg binary. Safe to run on
Windows, outside Docker, without any models installed.
Current unit files:
- `test_video_config.py` — `VideoConfig.from_dict` round-trip, LoRA target validation
- `test_video_engine_logic.py` — prompt derivation, library cursor, frame fitting
- `test_pipeline_video_branch.py` — pipeline takes the video path iff engine is ready
- `test_musetalk_fit_frames.py` — frame-length adjustment to match audio duration
- `test_muxer_ffmpeg.py` — ffmpeg command construction
## Component tests — slow, GPU-required, run inside Docker
Each script in `tests/component/` exercises one subsystem end-to-end against
the real models. They are ordered to match the implementation phases:
Each script in `tests/component/` exercises one subsystem end-to-end
against the real models. The numbered prefix reflects the implementation
phase each script gates, and also serves as a reasonable run order when
debugging a fresh environment:
| Script | Phase | Tests |
|---|---|---|
| `test_01_video_skeleton.py` | 1 | VideoEngine loads, config gate respected |
| `test_02_wan22_loras.py` | 2 | Wan2.2 pipeline loads, LoRA stack applies |
| `test_03_idle_clip.py` | 3 | set_avatar → idle MP4, written to disk for eyeballing |
| `test_03_idle_clip.py` | 3 | `set_avatar` → idle MP4, written to disk for eyeballing |
| `test_04_library_prebake.py` | 4 | library mode pre-bakes N base clips |
| `test_05_musetalk_lipsync.py` | 5 | MuseTalk lip-sync on library frames + ffmpeg mux |
| `test_06_reflective.py` | 6 | reflective mode: fresh Wan2.2 per reply |
| `test_07_endpoints.py` | 7 | HTTP endpoints return sane responses |
| `test_08_lora_reload.py` | 8 | /api/reload-loras swaps LoRAs live |
| `test_08_lora_reload.py` | 8 | `/api/reload-loras` swaps LoRAs live |
| `test_09_gguf_generate.py` | 9 | GGUF-quantised DIT end-to-end I2V generation |
| `test_10_t5_encode.py` | 10 | T5 encoder (optionally fp8-quantised) on CUDA |
| `test_11_image_encode.py` | 11 | Avatar image → VAE latent path |
| `test_12_dit_single_step.py` | 12 | Single DIT step on the loaded expert(s) |
| `test_13_vae_decode.py` | 13 | VAE decode back to RGB frames |
Tests 09-13 are focused on the GGUF + Blackwell (SM120) path and are how
new quant schemes / attention backends get validated before wiring them
into the full pipeline.
Run one:
@@ -36,7 +56,7 @@ Run one:
docker compose exec voice-chat python -m tests.component.test_03_idle_clip
```
Run all (slow, ~20+ minutes on 5090):
Run all (slow, ~20+ minutes on a 5090):
```
docker compose exec voice-chat python -m tests.component.run_all
tests/component/test_02_wan22_loras.py
+23 -39
View File
@@ -1,26 +1,26 @@
"""Phase 2 component test: Wan2.2 pipeline + LoRA stacking.
"""Phase 2 component test: dense Wan2.2-TI2V-5B-Turbo pipeline + LoRA stacking.
Verifies:
- ``Wan22Pipeline`` loads successfully (exercises the real LightX2V
set_config -> init_runner flow).
- ``load_loras`` / ``unload_loras`` survive with the two user LoRAs at
``/cache/loras/wan22-[HL]-e8.safetensors``.
- ``load_loras`` / ``unload_loras`` survive with any user LoRAs at
``/cache/loras/*.safetensors`` (target='both', dense single DIT).
Supports both fp8 and GGUF DIT quantisation. Set the ``DIT_QUANT``
environment variable to switch (default: ``fp8-sgl``).
Supports any GGUF quant published in hum-ma/Wan2.2-TI2V-5B-Turbo-GGUF.
Set ``DIT_QUANT`` to switch (default: ``gguf-Q8_0``).
DIT_QUANT=gguf-Q4_K_M docker compose exec voice-chat \
python -m tests.component.test_02_wan22_loras
Requires GPU and a first-run download of both HF repos (base support files
~12 GB, DIT size depends on quant — fp8 ~30 GB, GGUF Q4_K_M ~19 GB).
Requires GPU and a first-run download of the base repo + GGUF DIT.
If LightX2V isn't installed the test is skipped.
Run (default fp8):
Run:
docker compose exec voice-chat python -m tests.component.test_02_wan22_loras
"""
from __future__ import annotations
import glob
import os
import sys
@@ -28,19 +28,9 @@ from tests.component._common import get_logger
log = get_logger("test_02")
# --- Quant-dependent defaults ------------------------------------------------
DIT_QUANT = os.environ.get("DIT_QUANT", "fp8-sgl")
if DIT_QUANT.startswith("gguf-"):
CONFIG_JSON = "/app/configs/lightx2v/wan22_i2v_gguf_distill.json"
DIT_REPO = "QuantStack/Wan2.2-I2V-A14B-GGUF"
else:
CONFIG_JSON = "/app/configs/lightx2v/wan22_i2v_fp8_distill.json"
DIT_REPO = "lightx2v/Wan2.2-Distill-Models"
LORA_HIGH = "/cache/loras/wan22-H-e8.safetensors"
LORA_LOW = "/cache/loras/wan22-L-e8.safetensors"
DIT_QUANT = os.environ.get("DIT_QUANT", "gguf-Q8_0")
CONFIG_JSON = "/app/configs/lightx2v/wan22_i2v_gguf_5b_turbo.json"
DIT_REPO = "hum-ma/Wan2.2-TI2V-5B-Turbo-GGUF"
def run():
@@ -57,13 +47,14 @@ def run():
"(quant=%s, dit_repo=%s)...", DIT_QUANT, DIT_REPO)
try:
pipe = Wan22Pipeline(
base_repo="Wan-AI/Wan2.2-I2V-A14B",
base_repo="Wan-AI/Wan2.2-TI2V-5B",
dit_repo=DIT_REPO,
config_json=CONFIG_JSON,
model_cls="wan2.2_moe_distill",
model_cls="wan2.2",
resolution=480,
fps=16,
dit_quant_scheme=DIT_QUANT,
t5_quantized=True,
)
except Exception as e:
log.error("FAIL: Wan22Pipeline construction raised: %s", e)
@@ -78,34 +69,27 @@ def run():
pipe.load_loras([])
log.info(" PASS")
if not (os.path.isfile(LORA_HIGH) and os.path.isfile(LORA_LOW)):
log.warning("SKIP: expected LoRA files not found at %s / %s",
LORA_HIGH, LORA_LOW)
lora_files = sorted(glob.glob("/cache/loras/*.safetensors"))
if not lora_files:
log.warning("SKIP: no LoRA files found in /cache/loras/")
log.info("ALL PASSED (partial — LoRA cases skipped)")
return
log.info("[case 3] load_loras with the two MoE distill LoRAs")
lora_path = lora_files[0]
log.info("[case 3] load_loras with one 5B-compatible LoRA (%s)", lora_path)
specs = [
LoRASpec(
path=LORA_HIGH,
path=lora_path,
weight=1.0,
target="high_noise",
name="wan22-H-e8",
),
LoRASpec(
path=LORA_LOW,
weight=1.0,
target="low_noise",
name="wan22-L-e8",
target="both",
name=os.path.basename(lora_path),
),
]
try:
pipe.load_loras(specs)
except Exception as e:
log.error("FAIL: load_loras raised: %s", e)
log.error("Check: switch_lora support for wan2.2_moe_distill in the "
"installed LightX2V build. If it errors there, pre-declare "
"LoRAs in the config_json 'lora_configs' field instead.")
log.error("Check: LoRA checkpoint shape matches dense 5B DIT.")
sys.exit(3)
log.info(" PASS: LoRAs applied")
tests/component/test_07_endpoints.py
+2 -2
View File
@@ -81,9 +81,9 @@ def run():
body = {
"loras": [
{"path": "/cache/loras/a.safetensors", "weight": 0.8,
"target": "high_noise", "name": "test-a"},
"target": "both", "name": "test-a"},
{"path": "/cache/loras/b.safetensors", "weight": 0.4,
"target": "low_noise"},
"target": "both"},
]
}
resp = client.post("/api/reload-loras", json=body)
tests/component/test_08_lora_reload.py
+9 -17
View File
@@ -32,28 +32,20 @@ def run():
write_bytes("phase8_idle_noloras.mp4", idle_a)
log.info("idle (no LoRAs) sha256=%s", hash_a[:16])
# Hot-reload with a distill LoRA
specs = [
LoRASpec(
path="lightx2v/Wan2.2-Distill-Loras:"
"wan2.2_i2v_A14b_high_noise_lora_rank64_lightx2v_4step.safetensors",
weight=1.0,
target="high_noise",
name="distill-hi",
),
]
engine.load_loras(specs)
# Hot-reload flow: unload (no-op), reload empty list, verify clip still generates.
# There are no published 5B-Turbo-compatible LoRAs yet; when one exists,
# construct a LoRASpec(path=..., target="both", weight=1.0) and compare hashes.
engine.load_loras([])
engine.set_avatar(avatar_path)
idle_b = engine.get_idle_clip()
assert idle_b is not None
hash_b = hashlib.sha256(idle_b).hexdigest()
write_bytes("phase8_idle_withlora.mp4", idle_b)
log.info("idle (with LoRA) sha256=%s", hash_b[:16])
write_bytes("phase8_idle_reloaded.mp4", idle_b)
log.info("idle (post-reload) sha256=%s", hash_b[:16])
if hash_a != hash_b:
log.info("PASS: idle clip changed after LoRA reload")
else:
log.warning("clips identical — LoRA may not be applied; eyeball _out/*.mp4")
log.info("PASS: hot-reload round-trip completed "
"(hash match=%s — expected without a real LoRA applied).",
hash_a == hash_b)
if __name__ == "__main__":
tests/component/test_09_gguf_generate.py
+7 -12
View File
@@ -1,10 +1,10 @@
"""Quick smoke test: generate a video clip with the GGUF pipeline.
"""Quick smoke test: generate a video clip with the dense 5B Turbo GGUF pipeline.
Calls Wan22Pipeline.generate_i2v directly (no MuseTalk, no VideoEngine)
and writes the result to tests/component/_out/phase9_gguf.mp4.
Run:
docker compose exec -e DIT_QUANT=gguf-Q4_K_M voice-chat \
docker compose exec -e DIT_QUANT=gguf-Q8_0 voice-chat \
python -m tests.component.test_09_gguf_generate
"""
from __future__ import annotations
@@ -16,14 +16,9 @@ from tests.component._common import ensure_sample_avatar, get_logger, write_byte
log = get_logger("test_09")
DIT_QUANT = os.environ.get("DIT_QUANT", "gguf-Q4_K_M")
if DIT_QUANT.startswith("gguf-"):
CONFIG_JSON = "/app/configs/lightx2v/wan22_i2v_gguf_distill.json"
DIT_REPO = "QuantStack/Wan2.2-I2V-A14B-GGUF"
else:
CONFIG_JSON = "/app/configs/lightx2v/wan22_i2v_fp8_distill.json"
DIT_REPO = "lightx2v/Wan2.2-Distill-Models"
DIT_QUANT = os.environ.get("DIT_QUANT", "gguf-Q8_0")
CONFIG_JSON = "/app/configs/lightx2v/wan22_i2v_gguf_5b_turbo.json"
DIT_REPO = "hum-ma/Wan2.2-TI2V-5B-Turbo-GGUF"
def run():
@@ -38,10 +33,10 @@ def run():
log.info("Building pipeline (quant=%s)...", DIT_QUANT)
pipe = Wan22Pipeline(
base_repo="Wan-AI/Wan2.2-I2V-A14B",
base_repo="Wan-AI/Wan2.2-TI2V-5B",
dit_repo=DIT_REPO,
config_json=CONFIG_JSON,
model_cls="wan2.2_moe_distill",
model_cls="wan2.2",
resolution=480,
fps=16,
dit_quant_scheme=DIT_QUANT,
tests/component/test_10_t5_encode.py
+5 -10
View File
@@ -17,14 +17,9 @@ from tests.component._common import get_logger
log = get_logger("test_10")
DIT_QUANT = os.environ.get("DIT_QUANT", "gguf-Q4_K_M")
if DIT_QUANT.startswith("gguf-"):
CONFIG_JSON = "/app/configs/lightx2v/wan22_i2v_gguf_distill.json"
DIT_REPO = "QuantStack/Wan2.2-I2V-A14B-GGUF"
else:
CONFIG_JSON = "/app/configs/lightx2v/wan22_i2v_fp8_distill.json"
DIT_REPO = "lightx2v/Wan2.2-Distill-Models"
DIT_QUANT = os.environ.get("DIT_QUANT", "gguf-Q8_0")
CONFIG_JSON = "/app/configs/lightx2v/wan22_i2v_gguf_5b_turbo.json"
DIT_REPO = "hum-ma/Wan2.2-TI2V-5B-Turbo-GGUF"
def run():
@@ -36,10 +31,10 @@ def run():
log.info("Building pipeline (quant=%s) — this loads T5 + DIT weights...", DIT_QUANT)
pipe = Wan22Pipeline(
base_repo="Wan-AI/Wan2.2-I2V-A14B",
base_repo="Wan-AI/Wan2.2-TI2V-5B",
dit_repo=DIT_REPO,
config_json=CONFIG_JSON,
model_cls="wan2.2_moe_distill",
model_cls="wan2.2",
resolution=480,
fps=16,
dit_quant_scheme=DIT_QUANT,
tests/component/test_11_image_encode.py
+5 -10
View File
@@ -22,14 +22,9 @@ from tests.component._common import ensure_sample_avatar, get_logger
log = get_logger("test_11")
DIT_QUANT = os.environ.get("DIT_QUANT", "gguf-Q4_K_M")
if DIT_QUANT.startswith("gguf-"):
CONFIG_JSON = "/app/configs/lightx2v/wan22_i2v_gguf_distill.json"
DIT_REPO = "QuantStack/Wan2.2-I2V-A14B-GGUF"
else:
CONFIG_JSON = "/app/configs/lightx2v/wan22_i2v_fp8_distill.json"
DIT_REPO = "lightx2v/Wan2.2-Distill-Models"
DIT_QUANT = os.environ.get("DIT_QUANT", "gguf-Q8_0")
CONFIG_JSON = "/app/configs/lightx2v/wan22_i2v_gguf_5b_turbo.json"
DIT_REPO = "hum-ma/Wan2.2-TI2V-5B-Turbo-GGUF"
def run():
@@ -44,10 +39,10 @@ def run():
log.info("Building pipeline (quant=%s)...", DIT_QUANT)
pipe = Wan22Pipeline(
base_repo="Wan-AI/Wan2.2-I2V-A14B",
base_repo="Wan-AI/Wan2.2-TI2V-5B",
dit_repo=DIT_REPO,
config_json=CONFIG_JSON,
model_cls="wan2.2_moe_distill",
model_cls="wan2.2",
resolution=480,
fps=16,
dit_quant_scheme=DIT_QUANT,
tests/component/test_12_dit_single_step.py
+5 -10
View File
@@ -20,14 +20,9 @@ from tests.component._common import ensure_sample_avatar, get_logger
log = get_logger("test_12")
DIT_QUANT = os.environ.get("DIT_QUANT", "gguf-Q4_K_M")
if DIT_QUANT.startswith("gguf-"):
CONFIG_JSON = "/app/configs/lightx2v/wan22_i2v_gguf_distill.json"
DIT_REPO = "QuantStack/Wan2.2-I2V-A14B-GGUF"
else:
CONFIG_JSON = "/app/configs/lightx2v/wan22_i2v_fp8_distill.json"
DIT_REPO = "lightx2v/Wan2.2-Distill-Models"
DIT_QUANT = os.environ.get("DIT_QUANT", "gguf-Q8_0")
CONFIG_JSON = "/app/configs/lightx2v/wan22_i2v_gguf_5b_turbo.json"
DIT_REPO = "hum-ma/Wan2.2-TI2V-5B-Turbo-GGUF"
def run():
@@ -42,10 +37,10 @@ def run():
log.info("Building pipeline (quant=%s)...", DIT_QUANT)
pipe = Wan22Pipeline(
base_repo="Wan-AI/Wan2.2-I2V-A14B",
base_repo="Wan-AI/Wan2.2-TI2V-5B",
dit_repo=DIT_REPO,
config_json=CONFIG_JSON,
model_cls="wan2.2_moe_distill",
model_cls="wan2.2",
resolution=480,
fps=16,
dit_quant_scheme=DIT_QUANT,
tests/component/test_13_vae_decode.py
+5 -10
View File
@@ -19,14 +19,9 @@ from tests.component._common import ensure_sample_avatar, get_logger, write_byte
log = get_logger("test_13")
DIT_QUANT = os.environ.get("DIT_QUANT", "gguf-Q4_K_M")
if DIT_QUANT.startswith("gguf-"):
CONFIG_JSON = "/app/configs/lightx2v/wan22_i2v_gguf_distill.json"
DIT_REPO = "QuantStack/Wan2.2-I2V-A14B-GGUF"
else:
CONFIG_JSON = "/app/configs/lightx2v/wan22_i2v_fp8_distill.json"
DIT_REPO = "lightx2v/Wan2.2-Distill-Models"
DIT_QUANT = os.environ.get("DIT_QUANT", "gguf-Q8_0")
CONFIG_JSON = "/app/configs/lightx2v/wan22_i2v_gguf_5b_turbo.json"
DIT_REPO = "hum-ma/Wan2.2-TI2V-5B-Turbo-GGUF"
def run():
@@ -41,10 +36,10 @@ def run():
log.info("Building pipeline (quant=%s)...", DIT_QUANT)
pipe = Wan22Pipeline(
base_repo="Wan-AI/Wan2.2-I2V-A14B",
base_repo="Wan-AI/Wan2.2-TI2V-5B",
dit_repo=DIT_REPO,
config_json=CONFIG_JSON,
model_cls="wan2.2_moe_distill",
model_cls="wan2.2",
resolution=480,
fps=16,
dit_quant_scheme=DIT_QUANT,
tests/unit/test_video_config.py
+31 -26
View File
@@ -64,32 +64,39 @@ def test_lora_parse_full():
{
"loras": [
{
"path": "/tmp/hi.safetensors",
"path": "/tmp/a.safetensors",
"weight": 0.7,
"target": "high_noise",
"name": "hi-noise-style",
"target": "both",
"name": "style-a",
},
{
"path": "/tmp/lo.safetensors",
"path": "/tmp/b.safetensors",
"weight": 0.4,
"target": "low_noise",
"name": "lo-noise-style",
"target": "both",
"name": "style-b",
},
]
}
)
assert len(cfg.loras) == 2
assert cfg.loras[0].target == "high_noise"
assert cfg.loras[0].name == "hi-noise-style"
assert cfg.loras[1].target == "low_noise"
assert cfg.loras[0].target == "both"
assert cfg.loras[0].name == "style-a"
assert cfg.loras[1].target == "both"
assert cfg.loras[1].weight == 0.4
def test_lora_invalid_target_falls_back_to_both():
def test_lora_legacy_moe_target_coerced_to_both():
"""Legacy MoE configs with target='high_noise'/'low_noise' get coerced."""
cfg = VideoConfig.from_dict(
{"loras": [{"path": "/tmp/x.safetensors", "target": "bogus"}]}
{
"loras": [
{"path": "/tmp/hi.safetensors", "target": "high_noise"},
{"path": "/tmp/lo.safetensors", "target": "low_noise"},
{"path": "/tmp/x.safetensors", "target": "bogus"},
]
}
)
assert cfg.loras[0].target == "both"
assert all(l.target == "both" for l in cfg.loras)
def test_lora_entries_without_path_are_dropped():
@@ -107,8 +114,8 @@ def test_models_section_override():
"wan22_base_repo": "/local/weights/wan22",
"wan22_dit_repo": "/local/weights/wan22-dit",
"wan22_dit_quant_scheme": "gguf-Q4_K_M",
"wan22_config_json": "/local/cfg/fp8.json",
"wan22_model_cls": "wan2.2_moe",
"wan22_config_json": "/local/cfg/turbo.json",
"wan22_model_cls": "wan2.2",
"musetalk_path": "/local/weights/musetalk",
}
}
@@ -116,18 +123,16 @@ def test_models_section_override():
assert cfg.wan22_base_repo == "/local/weights/wan22"
assert cfg.wan22_dit_repo == "/local/weights/wan22-dit"
assert cfg.wan22_dit_quant_scheme == "gguf-Q4_K_M"
assert cfg.wan22_config_json == "/local/cfg/fp8.json"
assert cfg.wan22_model_cls == "wan2.2_moe"
assert cfg.wan22_config_json == "/local/cfg/turbo.json"
assert cfg.wan22_model_cls == "wan2.2"
assert cfg.musetalk_model_path == "/local/weights/musetalk"
def test_models_section_backwards_compat_fp8_repo():
"""Old config key wan22_fp8_repo still works via fallback."""
cfg = VideoConfig.from_dict(
{
"models": {
"wan22_fp8_repo": "/local/weights/wan22-fp8",
}
}
)
assert cfg.wan22_dit_repo == "/local/weights/wan22-fp8"
def test_models_section_defaults_to_5b_turbo():
cfg = VideoConfig.from_dict({})
assert cfg.wan22_base_repo == "Wan-AI/Wan2.2-TI2V-5B"
assert cfg.wan22_dit_repo == "hum-ma/Wan2.2-TI2V-5B-Turbo-GGUF"
assert cfg.wan22_dit_quant_scheme == "gguf-Q8_0"
assert cfg.wan22_t5_quantized is True
assert cfg.wan22_model_cls == "wan2.2"
assert cfg.wan22_config_json == "/app/configs/lightx2v/wan22_i2v_gguf_5b_turbo.json"