tuieval 0.2.0.dev3__tar.gz → 0.2.0.dev4__tar.gz

{tuieval-0.2.0.dev3 → tuieval-0.2.0.dev4}/CHANGELOG.md

@@ -2,6 +2,7 @@
 
 ## Unreleased
 
+- Warm-start tuning recognises fine-tunes whose GGUF describes the MTP draft layer differently (e.g. listing KV heads per layer) as the same model shape, so they start from an already-tuned sibling instead of tuning in full.
 - The fit check (context sized from the GGUF header, llama.cpp's memory use) only applies to servers whose command takes `{ctx}`. Servers that size their own memory keep the model's `max_context` instead of an estimate that didn't apply to them; `fit_check = true|false` on a server overrides it.
 - `tuieval export pi` has no default presets path any more: set `[export.pi] presets` to your llama.cpp router's `--models-preset` file.
 - A new README screenshot from a demo workspace.

{tuieval-0.2.0.dev3 → tuieval-0.2.0.dev4}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.5
 Name: tuieval
-Version: 0.2.0.dev3
+Version: 0.2.0.dev4
 Summary: Evaluate local and frontier LLMs on your own questions: accuracy, speed, tokens and PASS/FAIL verdicts, in the terminal.
 Project-URL: Homepage, https://github.com/ashe-wb/tuieval
 Project-URL: Issues, https://github.com/ashe-wb/tuieval/issues

{tuieval-0.2.0.dev3 → tuieval-0.2.0.dev4}/src/tuieval/_version.py

@@ -18,7 +18,7 @@ version_tuple: tuple[int | str, ...]
 commit_id: str | None
 __commit_id__: str | None
 
-__version__ = version = '0.2.0.dev3'
-__version_tuple__ = version_tuple = (0, 2, 0, 'dev3')
+__version__ = version = '0.2.0.dev4'
+__version_tuple__ = version_tuple = (0, 2, 0, 'dev4')
 
 __commit_id__ = commit_id = None

{tuieval-0.2.0.dev3 → tuieval-0.2.0.dev4}/src/tuieval/tune.py

@@ -315,8 +315,11 @@ def family(path):
         i = machines.read_gguf(engine_mod.expand(path))
     except (OSError, ValueError):
         return None
+    # The MTP (nextn) draft layers at the end are left out: GGUFs of the same model describe them
+    # differently, and whether draft-mtp is offered is decided per model (knobs) anyway.
+    heads = i["kv_heads_per_layer"][:len(i["kv_heads_per_layer"]) - i["mtp_layers"]]
     return (i["architecture"], i["layers"], i["embedding"], i["experts"], i["experts_used"],
-            i["head_dim_k"], i["head_dim_v"], tuple(i["kv_heads_per_layer"]))
+            i["head_dim_k"], i["head_dim_v"], tuple(heads))
 
 
 @dataclasses.dataclass

{tuieval-0.2.0.dev3 → tuieval-0.2.0.dev4}/tests/test_tuieval.py

@@ -225,17 +225,25 @@ class Units(unittest.TestCase):
             self.assertEqual(p.modules, ["some_missing_module"])
 
 
-def fake_gguf(path, embedding=5120, size=0):
-    """A GGUF header with just the keys machines.read_gguf reads, padded to size bytes."""
+def fake_gguf(path, embedding=5120, size=0, kv_heads=4):
+    """A GGUF header with just the keys machines.read_gguf reads, padded to size bytes. kv_heads: a
+    number, or one per layer (a list, as some GGUFs store it)."""
     import struct
     s = lambda t: struct.pack("<Q", len(t)) + t.encode()  # noqa: E731
     kv = [("general.architecture", 8, "qwen35"), ("qwen35.block_count", 4, 64),
           ("qwen35.embedding_length", 4, embedding), ("qwen35.attention.head_count", 4, 24),
-          ("qwen35.attention.head_count_kv", 4, 4), ("qwen35.attention.key_length", 4, 256),
+          ("qwen35.attention.head_count_kv", 9 if isinstance(kv_heads, list) else 4, kv_heads),
+          ("qwen35.attention.key_length", 4, 256),
           ("qwen35.full_attention_interval", 4, 4), ("qwen35.nextn_predict_layers", 4, 1)]
     out = b"GGUF" + struct.pack("<IQQ", 3, 0, len(kv))
     for key, t, v in kv:
-        out += s(key) + struct.pack("<I", t) + (s(v) if t == 8 else struct.pack("<I", v))
+        out += s(key) + struct.pack("<I", t)
+        if t == 8:
+            out += s(v)
+        elif t == 9:   # an array of uint32
+            out += struct.pack("<IQ", 4, len(v)) + b"".join(struct.pack("<I", x) for x in v)
+        else:
+            out += struct.pack("<I", v)
     with open(path, "wb") as f:
         f.write(out + b"\0" * max(0, size - len(out)))
 
@@ -316,6 +324,15 @@ class WarmTune(unittest.TestCase):
         self.tune.measure = self.orig
         self.tmp.cleanup()
 
+    def test_family_ignores_the_mtp_layer(self):
+        # one GGUF lists KV heads per layer, none on the MTP (last) layer; the other gives one number
+        d = self.tmp.name
+        per_layer = [4 if (i + 1) % 4 == 0 else 0 for i in range(63)] + [0]
+        fake_gguf(f"{d}/listed.gguf", kv_heads=per_layer)
+        self.assertEqual(self.tune.family(f"{d}/listed.gguf"), self.tune.family(f"{d}/new.gguf"))
+        fake_gguf(f"{d}/other-shape.gguf", kv_heads=[8 if (i + 1) % 4 == 0 else 0 for i in range(64)])
+        self.assertNotEqual(self.tune.family(f"{d}/other-shape.gguf"), self.tune.family(f"{d}/new.gguf"))
+
     def test_option_index(self):
         opts = self.KNOBS["spec"]
         self.assertEqual(self.tune.option_index(opts, ["-t", "6", "--spec-type", "draft-mtp"]), 2)