d3-flash v3.0.2

MODEL_MANIFEST.json

@@ -557,8 +557,8 @@
  },
  "runtime": {
   "files_sha256": {
-   "d3_runtime.py": "88fcaf3451edcc2139ffcb0db53ea4f9d03152e2736e0ffbb47386b40f618d52",
-   "d3_fast.py": "c8f6e74968e2835732e0865814b98a21e4b604c80ecb0416c270219eb8e64c05",
+   "d3_runtime.py": "139ece85007c5771b5e374da9658dc8832eb23aabfeaabe29f18e56d59d7f82f",
+   "d3_fast.py": "68a1fb9a2c3bde9ac2763bf504decae592ca12722698fbde1b39a31262955535",
    "d3_kernels.py": "8f7534ca1e1846ebf7c4c73058002f3be4dbf0d5d426f0815150ae0cbe7677bd",
    "modeling_d3.py": "9a718530f18d33d706a44d79be463ff43963fa3a9f808ce934a6489fd09042aa",
    "pipeline_d3.py": "c0975a91316bcda9c09c5f14bf634ceec94e6980ed7e76a29b718fb451fe4091",
@@ -581,11 +581,11 @@
    }
   ]
  },
- "built_utc": "2026-10-10T21:45:40+00:00",
+ "built_utc": "2026-10-11T00:16:22+00:00",
  "files_sha256": {
   "LICENSE": "c71d239df91726fc519c6eb72d318ec65820627232b2f796219e87dcf35d0ab4",
   "NOTICE": "ae7ffb4d788c076d655394f0f116ad4b7de54e3bca9b655df5c669ea5bae74f6",
-  "README.md": "4aeebb539c454a6c663b887b9bff76ec7107ae711ac03f3778a0f2156bb89b09",
+  "README.md": "9b7836b3165c07291603c1a61c179dc0c6a40bbe495e4fdedb629a55f882edd2",
   "assets/banner.png": "d621431eb5d1b20650a35a2e27c30f372631592cabff260679f12be9813eebc5",
   "assets/example-receipt.png": "b9dbb8b103bacd45f1de824d844a3c07b81d57cb1a7466712ef060ad7dcb3926",
   "assets/index-areas.png": "cea16dca531b6b9bb039e3199061e22aefc619ec7a3e5d5455f00bd20ae25e6d",
@@ -593,10 +593,10 @@
   "chat_template.jinja": "a4aee8afcf2e0711942cf848899be66016f8d14a889ff9ede07bca099c28f715",
   "config.json": "9c3e6eea217964096c7b032285bab6757720a847fbb9d0a04e551d2b1606c0eb",
   "d3_engine.py": "ab638d785325d944d3dac891f77448dfc91cad1b8d55a7302bd527c5f1e0e251",
-  "d3_fast.py": "c8f6e74968e2835732e0865814b98a21e4b604c80ecb0416c270219eb8e64c05",
+  "d3_fast.py": "68a1fb9a2c3bde9ac2763bf504decae592ca12722698fbde1b39a31262955535",
   "d3_format.py": "187a0f359f6f89a844e723377769556f78133b118851e702e4b0fe5123ccb0f1",
   "d3_kernels.py": "8f7534ca1e1846ebf7c4c73058002f3be4dbf0d5d426f0815150ae0cbe7677bd",
-  "d3_runtime.py": "88fcaf3451edcc2139ffcb0db53ea4f9d03152e2736e0ffbb47386b40f618d52",
+  "d3_runtime.py": "139ece85007c5771b5e374da9658dc8832eb23aabfeaabe29f18e56d59d7f82f",
   "d3_server.py": "11211faa83fee50b957903f0d596c153b0921cb5414f18edee31d8dac76db9fc",
   "decision_config.json": "0269c72ac055d8149a7a778c477780af361f4c1cde3fd0e5ba9af791453ece18",
   "merges.txt": "a9d356d7bdf1ef4949e3e748e95b8e10ad9d4e2e838eddc38a0a7b6b94d1db8d",
@@ -626,10 +626,10 @@
   "chat_template.jinja": 7756,
   "config.json": 3093,
   "d3_engine.py": 3726,
-  "d3_fast.py": 21284,
+  "d3_fast.py": 22331,
   "d3_format.py": 4881,
   "d3_kernels.py": 16129,
-  "d3_runtime.py": 57795,
+  "d3_runtime.py": 59432,
   "d3_server.py": 8333,
   "decision_config.json": 5502,
   "merges.txt": 3353259,

README.md

@@ -31,7 +31,7 @@ tags:
 - **Jev Decision Index 0.3, public suite: 57.79**, measured with the official 0.3 kit on the released weights: all 140,178 public requests answered, none unsupported.
 - **+11.0 on the public suite over Decision 2.0** (its 9B model: 46.76 on the board), ahead in all five areas.
 - **Reads images:** multiple images per request (PNG, JPEG or WebP), given as paths, URLs, PIL images or base64 data URLs; every question of the request sees all of them.
-- **Speed:** a median of 19.4 ms for a text request and 151.9 ms for a request with an image, on one AMD Instinct MI325X GPU, one request at a time.
+- **Speed:** a median of 19.1 ms for a text request and 149.4 ms for a request with an image, on one AMD Instinct MI325X GPU, one request at a time.
 - **Many questions, one call:** Choice, Yes / No and Score questions about the same input are answered together, each from its own forward pass over the input, with a probability for every option.
 
 ## Quickstart

d3_fast.py

@@ -71,6 +71,30 @@ def fusable(model: Any) -> str | None:
     return "needs " + ", ".join(missing) if missing else None
 
 
+def kernels_next_to_this_file():
+    """``d3_kernels.py`` of this file's directory, imported by path once per file (see ``d3_runtime.fast_module``)."""
+    import hashlib
+    import importlib.util
+    import sys
+    from pathlib import Path
+
+    # Not resolved: in a Hub snapshot the .py names are links to blobs without a suffix.
+    path = Path(__file__).absolute().with_name("d3_kernels.py")
+    name = f"d3_kernels_{hashlib.sha256(str(path).encode()).hexdigest()[:16]}"
+    if name not in sys.modules:
+        spec = importlib.util.spec_from_file_location(name, str(path))
+        if spec is None or spec.loader is None:
+            raise ImportError(f"no loader for {path}")
+        module = importlib.util.module_from_spec(spec)
+        sys.modules[name] = module
+        try:
+            spec.loader.exec_module(module)
+        except BaseException:
+            del sys.modules[name]
+            raise
+    return sys.modules[name]
+
+
 class FusedLayers:
     """The decoder layers and final norm of a Qwen3.5 text model through ``d3_kernels`` (eager values)."""
 
@@ -82,7 +106,10 @@ class FusedLayers:
         try:
             from . import d3_kernels as kernels
         except ImportError:
-            kernels = importlib.import_module("d3_kernels")
+            try:
+                kernels = importlib.import_module("d3_kernels")
+            except ImportError:
+                kernels = kernels_next_to_this_file()
         self.k = kernels
         self.torch = torch
         self.chunk = modeling.torch_chunk_gated_delta_rule

d3_runtime.py

@@ -463,6 +463,39 @@ def sha256_file(path: Path) -> str:
     return digest.hexdigest()
 
 
+def load_file_module(path: Path):
+    """Import a flat package file by its path, once per file (two checkpoints can share a process)."""
+    import importlib.util
+    import sys
+
+    # Not resolved: in a Hub snapshot the .py name is a link to a blob without a suffix, which has no loader.
+    path = Path(path).absolute()
+    name = f"{path.stem}_{hashlib.sha256(str(path).encode()).hexdigest()[:16]}"
+    if name not in sys.modules:
+        spec = importlib.util.spec_from_file_location(name, str(path))
+        if spec is None or spec.loader is None:
+            raise ImportError(f"no loader for {path}")
+        module = importlib.util.module_from_spec(spec)
+        sys.modules[name] = module
+        try:
+            spec.loader.exec_module(module)
+        except BaseException:
+            del sys.modules[name]
+            raise
+    return sys.modules[name]
+
+
+def fast_module(root: Path):
+    """``d3_fast.py`` of the checkpoint directory when it is not importable as a module.
+
+    ``AutoModel.from_pretrained(..., trust_remote_code=True)`` copies only the modeling files and their
+    ``from .x import y`` imports into its module cache; the package's own copy (verified against
+    ``MODEL_MANIFEST.json``) is next to the weights.
+    """
+    path = Path(root) / "d3_fast.py"
+    return load_file_module(path) if path.is_file() else None
+
+
 def resolve_dir(
     name_or_path: str | os.PathLike, revision: str | None = None, **hub: Any
 ) -> Path:
@@ -799,8 +832,14 @@ class D3:
                 self.image_unavailable = (
                     f"the image processor failed to load ({type(exc).__name__}: {exc})"
                 )
-        self.fast_skipped = None if d3_fast else "d3_fast.py is not present"
-        self.fast = d3_fast.install(self) if d3_fast else None
+        fast, reason = d3_fast, "d3_fast.py is not present"
+        if fast is None:
+            try:
+                fast = fast_module(self.root)
+            except Exception as exc:  # noqa: BLE001 - the plain path always loads
+                reason = f"d3_fast.py failed to import ({type(exc).__name__}: {exc})"
+        self.fast_skipped = None if fast else reason
+        self.fast = fast.install(self) if fast else None
         self.loaded_seconds = time.perf_counter() - started
 
     @classmethod