transformersを4.x系に固定し、text encoder付きLoDA読み込み時の

IndexErrorを解消。あわせて外部VAEロードの既存バグも修正した。

問題:
kohya形式LoRA（pixelartV3, pixel_f2, pixel sprites等）のロード時に
Diffusers内部でIndexError: list index out of rangeが発生し、サーバが
起動失敗していた。LCM LoRAはtext encoderキーを含まないため成功していた。

原因:
Transformers 5.x系のCLIPTextModel.named_modules()が返すモジュール名
（encoder.layers.0.self_attn.k_proj）と、Diffusersがkohya形式LoRAを
PEFT形式へ変換後のキー（text_model.encoder.layers.0...lora_B.weight）の
間でtext_model.プレフィックスの有無が不一致となり、rank_dictが空になって
get_peft_kwargsでlist(rank_dict.values())[0]がIndexErrorになる。
Transformers 4.x系ではnamed_modules()がtext_model.プレフィックス付きの
モジュール名を返すため、PEFTキーと一致し正常にロードできる。

対応:
transformers>=4.45,<5 に固定。4.x系最新版の4.57.6を採用。
これに伴いdiffusers 0.40.0→0.39.0、huggingface-hub 1.31.0→0.36.2、
tokenizers 0.23.2→0.22.2に変更された。peft 0.20.0は変更なし。
LoDAロード処理のコードは変更せず、標準APIのpipeline.load_lora_weights()
をそのまま使用。モンキーパッチやキー書き換え等の互換性パッチは追加していない。

あわせて外部VAEロードの既存バグも修正:
AutoencoderKL.from_single_fileがネットワークからconfig.jsonを取得
しようとして失敗する問題を、_vae_config_dir()でローカルhuggingface
cache内のvae/config.jsonディレクトリを動的に見つけてconfigパラメータ
に渡すことで解消。

変更内容:
- pyproject.toml: transformers>=4.45 を transformers>=4.45,<5 に変更
- uv.lock: transformers 5.17.0→4.57.6, diffusers 0.40.0→0.39.0,
  huggingface-hub 1.31.0→0.36.2, tokenizers 0.23.2→0.22.2
- pipeline_factory.py: _vae_config_dir()メソッド追加、VAEロード時に
  configパラメータでローカルconfigディレクトリを指定
- tests/test_pipeline_factory.py: _vae_config_dirをモンキーパッチ、
  VAE config引数の検証追加
- script/smoke_matrix.py: 10パターン網羅検証スクリプト追加
  (4 checkpoint × VAE × 4種LoRA × scheduler)

検証:
- ruff format/check、pyright 0 errors、pytest 61 passed
- 実GPU・実モデルで10パターン全て成功（10 ok / 0 fail）
  - 4種checkpoint単体: 全て成功
  - 外部VAE差し替え: 成功
  - pixel sprites / pixelartV3 / pixel_f2: 全て成功（IndexError解消）
  - LCM LoRA + lcm scheduler: 成功
  - animekawa + 外部VAE + pixelartV3 + euler_a: 成功

1 files changed, 304 insertions, 0 deletions

diff --git a/script/smoke_matrix.py b/script/smoke_matrix.py
new file mode 100644
index 0000000..7fc3ec2
--- /dev/null
+++ b/script/smoke_matrix.py
@@ -0,0 +1,304 @@
+"""Run generate_image across all checkpoint/VAE/LoRA/scheduler combinations.
+
+For each pattern, this script writes a temporary TOML config, starts the
+``mcp-sd15`` server with ``streamable-http`` transport, calls ``generate_image``
+through the HTTP endpoint, saves the PNG, and records the result.
+
+Usage::
+
+    uv run python script/smoke_matrix.py --output-dir data/outputs/matrix
+"""
+
+from __future__ import annotations
+
+import argparse
+import asyncio
+import base64
+import json
+import subprocess
+import sys
+import time
+from pathlib import Path
+
+from mcp import Client
+from mcp.types import ImageContent
+
+DEFAULT_PROMPT = "a small red cube on a white table, studio photograph, detailed"
+DEFAULT_NEGATIVE = "blurry, low quality"
+
+PATTERNS: list[dict[str, object]] = [
+    {
+        "name": "01_v15_base",
+        "checkpoint": "v1-5-pruned-emaonly.safetensors",
+        "vae": None,
+        "loras": [],
+        "scheduler": "dpmpp_2m_karras",
+        "steps": 20,
+        "cfg_scale": 7.0,
+        "prompt": DEFAULT_PROMPT,
+        "negative": DEFAULT_NEGATIVE,
+    },
+    {
+        "name": "02_animekawa_base",
+        "checkpoint": "animekawa_v10.safetensors",
+        "vae": None,
+        "loras": [],
+        "scheduler": "dpmpp_2m_karras",
+        "steps": 20,
+        "cfg_scale": 7.0,
+        "prompt": DEFAULT_PROMPT,
+        "negative": DEFAULT_NEGATIVE,
+    },
+    {
+        "name": "03_amore_base",
+        "checkpoint": "animeModel_amore.safetensors",
+        "vae": None,
+        "loras": [],
+        "scheduler": "dpmpp_2m_karras",
+        "steps": 20,
+        "cfg_scale": 7.0,
+        "prompt": DEFAULT_PROMPT,
+        "negative": DEFAULT_NEGATIVE,
+    },
+    {
+        "name": "04_animepastel_base",
+        "checkpoint": "animePastelDream_softBakedVae_full_fp16.safetensors",
+        "vae": None,
+        "loras": [],
+        "scheduler": "dpmpp_2m_karras",
+        "steps": 20,
+        "cfg_scale": 7.0,
+        "prompt": DEFAULT_PROMPT,
+        "negative": DEFAULT_NEGATIVE,
+    },
+    {
+        "name": "05_v15_external_vae",
+        "checkpoint": "v1-5-pruned-emaonly.safetensors",
+        "vae": "vaeKlF8Anime2_klF8Anime2VAE.safetensors",
+        "loras": [],
+        "scheduler": "dpmpp_2m_karras",
+        "steps": 20,
+        "cfg_scale": 7.0,
+        "prompt": DEFAULT_PROMPT,
+        "negative": DEFAULT_NEGATIVE,
+    },
+    {
+        "name": "06_v15_lora_pixel_sprites",
+        "checkpoint": "v1-5-pruned-emaonly.safetensors",
+        "vae": None,
+        "loras": [{"file": "pixel sprites.safetensors", "weight": 0.8}],
+        "scheduler": "dpmpp_2m_karras",
+        "steps": 20,
+        "cfg_scale": 7.0,
+        "prompt": "pixel art, a small red cube on a white table, detailed",
+        "negative": DEFAULT_NEGATIVE,
+    },
+    {
+        "name": "07_v15_lora_pixelartV3",
+        "checkpoint": "v1-5-pruned-emaonly.safetensors",
+        "vae": None,
+        "loras": [{"file": "pixelartV3.safetensors", "weight": 0.8}],
+        "scheduler": "dpmpp_2m_karras",
+        "steps": 20,
+        "cfg_scale": 7.0,
+        "prompt": "pixelart, a small red cube on a white table, detailed",
+        "negative": DEFAULT_NEGATIVE,
+    },
+    {
+        "name": "08_v15_lora_pixel_f2",
+        "checkpoint": "v1-5-pruned-emaonly.safetensors",
+        "vae": None,
+        "loras": [{"file": "pixel_f2.safetensors", "weight": 0.8}],
+        "scheduler": "dpmpp_2m_karras",
+        "steps": 20,
+        "cfg_scale": 7.0,
+        "prompt": "pixel, a small red cube on a white table, detailed",
+        "negative": DEFAULT_NEGATIVE,
+    },
+    {
+        "name": "09_v15_lcm_lora",
+        "checkpoint": "v1-5-pruned-emaonly.safetensors",
+        "vae": None,
+        "loras": [{"file": "LCM_LoRA_Weights_SD15.safetensors", "weight": 1.0}],
+        "scheduler": "lcm",
+        "steps": 4,
+        "cfg_scale": 1.0,
+        "prompt": DEFAULT_PROMPT,
+        "negative": DEFAULT_NEGATIVE,
+    },
+    {
+        "name": "10_animekawa_vae_lora",
+        "checkpoint": "animekawa_v10.safetensors",
+        "vae": "vaeKlF8Anime2_klF8Anime2VAE.safetensors",
+        "loras": [{"file": "pixelartV3.safetensors", "weight": 0.7}],
+        "scheduler": "euler_a",
+        "steps": 20,
+        "cfg_scale": 7.0,
+        "prompt": "pixelart, a small red cube on a white table, detailed",
+        "negative": DEFAULT_NEGATIVE,
+    },
+]
+
+
+def make_toml(pattern: dict[str, object], port: int) -> str:
+    lines: list[str] = []
+    lines.append("[model]")
+    lines.append('root = "."')
+    lines.append(f'checkpoint = "{pattern["checkpoint"]}"')
+    if pattern["vae"]:
+        lines.append(f'vae = "{pattern["vae"]}"')
+    for lora in pattern["loras"]:  # type: ignore[union-attr]
+        lines.append("[[model.lora]]")
+        lines.append(f'file = "{lora["file"]}"')  # type: ignore[index]
+        lines.append(f"weight = {lora['weight']}")  # type: ignore[index]
+    lines.append("[generation]")
+    lines.append(f"steps = {pattern['steps']}")
+    lines.append(f"cfg_scale = {pattern['cfg_scale']}")
+    lines.append(f'scheduler = "{pattern["scheduler"]}"')
+    lines.append('default_size = "square_512"')
+    lines.append("[output]")
+    lines.append("enabled = false")
+    lines.append("[server]")
+    lines.append('transport = "streamable-http"')
+    lines.append('host = "127.0.0.1"')
+    lines.append(f"port = {port}")
+    return "\n".join(lines) + "\n"
+
+
+async def run_one(
+    pattern: dict[str, object],
+    config_path: Path,
+    output_path: Path,
+    port: int,
+) -> dict[str, object]:
+    toml = make_toml(pattern, port)
+    config_path.write_text(toml, encoding="utf-8")
+
+    proc = subprocess.Popen(
+        [sys.executable, "-m", "mcp_sd15", "--config", str(config_path)],
+        stdout=subprocess.PIPE,
+        stderr=subprocess.STDOUT,
+        text=True,
+        bufsize=1,
+    )
+
+    result: dict[str, object] = {"name": pattern["name"], "status": "unknown"}
+    started = time.monotonic()
+
+    try:
+        # Wait for server to be ready (up to 120s for model loading)
+        url = f"http://127.0.0.1:{port}/mcp"
+        for _wait in range(120):
+            if proc.poll() is not None:
+                # Process died; drain stdout
+                dead_output = ""
+                if proc.stdout:
+                    dead_output = proc.stdout.read()
+                result["status"] = "server_died"
+                result["output"] = dead_output[-3000:]
+                return result
+            try:
+                import httpx
+
+                resp = httpx.get(url, timeout=2.0)
+                if resp.status_code == 400:
+                    break
+            except Exception:
+                pass
+            await asyncio.sleep(1)
+        else:
+            result["status"] = "server_timeout"
+            return result
+
+        async with Client(url, raise_exceptions=True) as client:
+            tools = (await client.list_tools()).tools
+            tool_names = {tool.name for tool in tools}
+            if tool_names != {"generate_image", "generate_guided_image", "inpaint_image"}:
+                result["status"] = "wrong_tools"
+                result["tools"] = list(tool_names)
+                return result
+
+            call_result = await client.call_tool(
+                "generate_image",
+                {
+                    "positive_prompt": pattern["prompt"],  # type: ignore[index]
+                    "negative_prompt": pattern["negative"],  # type: ignore[index]
+                    "size": "square_512",
+                },
+            )
+
+        if call_result.is_error or len(call_result.content) != 1:
+            result["status"] = "tool_error"
+            result["detail"] = str(call_result)
+            return result
+
+        content = call_result.content[0]
+        if not isinstance(content, ImageContent) or content.mime_type != "image/png":
+            result["status"] = "wrong_content"
+            result["detail"] = str(content)
+            return result
+
+        output_path.parent.mkdir(parents=True, exist_ok=True)
+        output_path.write_bytes(base64.b64decode(content.data, validate=True))
+        elapsed = round(time.monotonic() - started, 1)
+        result["status"] = "ok"
+        result["elapsed"] = elapsed
+        result["output_file"] = str(output_path)
+        return result
+    except Exception as exc:
+        result["status"] = "exception"
+        result["detail"] = str(exc)
+        return result
+    finally:
+        proc.terminate()
+        try:
+            proc.wait(timeout=10)
+        except subprocess.TimeoutExpired:
+            proc.kill()
+
+
+async def main(args: argparse.Namespace) -> None:
+    config_path = args.data_dir / "config.matrix.toml"
+    results: list[dict[str, object]] = []
+
+    for i, pattern in enumerate(PATTERNS):
+        name = pattern["name"]
+        print(f"[{i + 1}/{len(PATTERNS)}] {name} ...", flush=True)
+        output_path = args.output_dir / f"{name}.png"
+        result = await run_one(pattern, config_path, output_path, args.port + i)
+        status = result["status"]
+        if status == "ok":
+            print(
+                f"  -> OK  elapsed={result.get('elapsed')}s  file={result.get('output_file')}",
+                flush=True,
+            )
+        else:
+            print(f"  -> FAIL  status={status}  detail={result.get('detail', '')}", flush=True)
+        results.append(result)
+
+    config_path.unlink(missing_ok=True)
+
+    summary_path = args.output_dir / "matrix_results.json"
+    summary_path.parent.mkdir(parents=True, exist_ok=True)
+    summary_path.write_text(json.dumps(results, indent=2, ensure_ascii=False), encoding="utf-8")
+
+    ok = sum(1 for r in results if r["status"] == "ok")
+    fail = len(results) - ok
+    print(f"\n=== Summary: {ok} ok / {fail} fail / {len(results)} total ===", flush=True)
+    if fail:
+        print("Failures:", flush=True)
+        for r in results:
+            if r["status"] != "ok":
+                print(f"  {r['name']}: {r['status']} {r.get('detail', '')}", flush=True)
+
+
+def parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--output-dir", type=Path, default=Path("data/outputs/matrix"))
+    parser.add_argument("--data-dir", type=Path, default=Path("data"))
+    parser.add_argument("--port", type=int, default=8130)
+    return parser.parse_args()
+
+
+if __name__ == "__main__":
+    asyncio.run(main(parse_args()))