From e029f42866d4c45e2062a22e84a8f4387bb68603 Mon Sep 17 00:00:00 2001 From: dualchimerra Date: Sat, 19 Sep 2026 21:21:38 +0200 Subject: [PATCH 1/3] fix(soup): close the mkstemp handle before writing the .part file On Windows the handle mkstemp leaves open blocks safetensors from writing the merged file ("access denied") and blocks the cleanup unlink, so every merge failed with a 500. Same fix for uploads. Co-Authored-By: Claude Opus 5 --- studio/services/soup.py | 15 +++++++++++++-- 1 file changed, 13 insertions(+), 2 deletions(-) diff --git a/studio/services/soup.py b/studio/services/soup.py index ac8bab5..753aeca 100644 --- a/studio/services/soup.py +++ b/studio/services/soup.py @@ -160,13 +160,24 @@ def _unlink(path: Path) -> None: raise SoupError(f"Could not delete the file: {exc}") from exc +def _temp_part(directory: Path) -> Path: + """Reserve a unique `.part` file and release its handle right away. + + mkstemp leaves the file open; on Windows that open handle blocks the + writer (safetensors gets "access denied") and the cleanup unlink. + """ + fd, name = tempfile.mkstemp(dir=str(directory), suffix=".part") + os.close(fd) + return Path(name) + + def save_upload(name: str, data: bytes) -> dict[str, Any]: """Store an uploaded adapter, after checking it is a readable safetensors.""" ensure_dirs() if not data: raise SoupError("The uploaded file is empty.") path = upload_path(name) - tmp = Path(tempfile.mkstemp(dir=str(UPLOAD_DIR), suffix=".part")[1]) + tmp = _temp_part(UPLOAD_DIR) try: tmp.write_bytes(data) # Parse before accepting: a truncated upload should fail here, not two @@ -409,7 +420,7 @@ def merge( meta = _soup_metadata(base_info, paths, raw_weights, weights, method) - tmp = Path(tempfile.mkstemp(dir=str(OUTPUT_DIR), suffix=".part")[1]) + tmp = _temp_part(OUTPUT_DIR) try: save_file(merged, str(tmp), metadata=meta) os.replace(tmp, out) From f58456bf7c890112ed843fd0148cac38482f7438 Mon Sep 17 00:00:00 2001 From: dualchimerra Date: Sat, 19 Sep 2026 21:21:39 +0200 Subject: [PATCH 2/3] fix(tunnel): wait for the CLI to exit before reading its exit code Output EOF can arrive a moment before Windows records the exit code, so poll() returned None ~30 ms after `tailscale funnel --bg` had already succeeded and the studio reported "did not finish". Co-Authored-By: Claude Opus 5 --- studio/services/tunnel.py | 8 ++++++-- 1 file changed, 6 insertions(+), 2 deletions(-) diff --git a/studio/services/tunnel.py b/studio/services/tunnel.py index 9de7b7d..785d077 100644 --- a/studio/services/tunnel.py +++ b/studio/services/tunnel.py @@ -463,8 +463,12 @@ def _run_until_exit(self, cmd: list[str]) -> tuple[list[str], Optional[int]]: waiting (killed after the timeout, e.g. Tailscale asking for consent).""" proc = self._spawn(cmd, dict(os.environ)) self._await(proc, lambda _line: None, until_exit=True) - code = proc.poll() - if code is None: + # Output EOF can arrive a moment before the OS records the exit code + # (seen on Windows: poll() still None ~30 ms after the CLI finished). + try: + code: Optional[int] = proc.wait(timeout=5) + except subprocess.TimeoutExpired: + code = None self._stop_process(proc) return list(self._log), code From 0556c881457124d8a34d8f73f4ba0076bbac1f31 Mon Sep 17 00:00:00 2001 From: dualchimerra Date: Sat, 19 Sep 2026 21:21:39 +0200 Subject: [PATCH 3/3] fix(anima): read the block count from the checkpoint; rebuild RoPE tables The loader hard-coded 28 blocks for every 2048-wide Anima, so the 40-block 2.9B checkpoint crashed on the saved pos_embedder.seq (256 vs 512) and, had it loaded, strict=False would have silently dropped 12 blocks. RoPE tables are derived from the config, so they are no longer read from the checkpoint. Co-Authored-By: Claude Opus 5 --- runtime/training/families/anima/loader.py | 12 ++++++++++++ 1 file changed, 12 insertions(+) diff --git a/runtime/training/families/anima/loader.py b/runtime/training/families/anima/loader.py index 7f646f7..064070f 100644 --- a/runtime/training/families/anima/loader.py +++ b/runtime/training/families/anima/loader.py @@ -188,6 +188,12 @@ def load_anima_model(transformer_path, device, dtype, repo_root, *, num_blocks, num_heads = 36, 40 else: raise RuntimeError(f"未知的 model_channels={model_channels}") + # 层数以 checkpoint 为准：同为 2048 宽也有 28 层（2B）和 40 层（2.9B）两种。 + # 写死 28 会让 strict=False 的加载静默丢掉多出来的层。 + header_blocks = block_count_from_header(transformer_path) + if header_blocks and header_blocks != num_blocks: + logger.info("Anima DiT 层数按 checkpoint 取 %d（默认 %d）", header_blocks, num_blocks) + num_blocks = header_blocks config = dict( max_img_h=1024, max_img_w=1024, max_frames=128, @@ -209,6 +215,12 @@ def load_anima_model(transformer_path, device, dtype, repo_root, *, # 加载权重 sd = _load_safetensors_state_dict(Path(transformer_path)) + # RoPE 表（seq = arange(最大长度) 等）由 config 推出；有的导出把它们按别的 + # 最大长度存进了权重（2.9B: seq 256 vs 模型 512），交给模型自己重建。 + sd = { + k: v for k, v in sd.items() + if not ("pos_embedder." in k and k.rsplit(".", 1)[-1] in ("seq", "dim_spatial_range", "dim_temporal_range")) + } info = _load_weights_best_effort(model, sd, label="Transformer") # 如果 checkpoint 中完全没有 llm_adapter 权重，随机初始化会把 cross-attn 条件搞乱，直接禁用更安全