Skip to content

Local UI to run and train LLMs and diffusion models. Supports GGUF, MLX, Qwen3.8, DeepSeek-V4.1, MiniMax-H3, Gemma 4, FLUX and more.

Studio: recognise model.safetensors-N-of-M shards for training resume (#15405) (#15408)

Indrajit swain4aeeb08682b1
diff --git a/studio/backend/core/training/provenance.py b/studio/backend/core/training/provenance.py
index 485d8cb172f..7161970eb06 100644
--- a/studio/backend/core/training/provenance.py
+++ b/studio/backend/core/training/provenance.py
@@ -24,7 +24,10 @@ _MODEL_LOAD_PREQUANTIZED_4BIT = "prequantized_4bit"
_MODEL_LOAD_RUNTIME_4BIT = "runtime_4bit"
_PREQUANTIZED_MODE_RE = re.compile(r"prequantized_(?:1[0-6]|[1-9])(?:_(?:1[0-6]|[1-9])){0,7}bit")
_REASON_RE = re.compile(r"[a-z0-9][a-z0-9_-]{0,63}")
+# unsloth/Qwen3.5-* shards are named ``model.safetensors-00001-of-00001.safetensors`` (#15405).
_MODEL_WEIGHT_RE = re.compile(
+ r"(?:"
+ r"model\.safetensors-\d+-of-\d+(?:\.safetensors)?|"
r"(?:"
r"model(?:-\d+-of-\d+)?|"
r"pytorch_model(?:-\d+-of-\d+)?|"
@@ -32,7 +35,8 @@ _MODEL_WEIGHT_RE = re.compile(
r"consolidated(?:[._-]\d+)?|"
r"mlx_model(?:-\d+-of-\d+)?|"
r"weights"
- r")\.(?:safetensors|bin|pt|pth|ckpt|npz)$",
+ r")\.(?:safetensors|bin|pt|pth|ckpt|npz)"
+ r")$",
re.IGNORECASE,
)
_DATASET_DATA_SUFFIXES = (
diff --git a/studio/backend/tests/test_training_provenance.py b/studio/backend/tests/test_training_provenance.py
index 326aa8298cf..ab57f9b319a 100644
--- a/studio/backend/tests/test_training_provenance.py
+++ b/studio/backend/tests/test_training_provenance.py
@@ -14,6 +14,8 @@ import pytest
from core.training.provenance import (
ExactResumeResourcesUnavailable,
RESOURCE_PROVENANCE_KEY,
+ _MODEL_WEIGHT_RE,
+ _snapshot_has_model_weights,
attest_loaded_dataset,
build_worker_provenance_event,
effective_training_load_in_4bit,
@@ -719,6 +721,24 @@ def test_config_only_snapshot_cannot_attest_model_weights(tmp_path):
)
+ "filename",
+ [
+ "model.safetensors-00001-of-00001.safetensors",
+ "model.safetensors-00001-of-00001",
+ "model-00000-of-00001.safetensors",
+ "model.safetensors",
+ ],
+)
+def test_extension_before_shard_weight_names_count_as_model_weights(tmp_path, filename):
+ """Qwen3.5 Hub shards use model.safetensors-00001-of-00001.safetensors (#15405)."""
+ assert _MODEL_WEIGHT_RE.fullmatch(filename)
+ snapshot = _model_snapshot(tmp_path, "unsloth/Qwen3.5-2B", "qwen35", weights = False)
+ (snapshot / filename).write_bytes(b"weights")
+ assert _snapshot_has_model_weights(snapshot)
+ assert exact_model_snapshot_path(str(snapshot), "unsloth/Qwen3.5-2B") == str(snapshot.resolve())
+
+
def test_exact_model_snapshot_accepts_own_blob_symlink(tmp_path):
snapshot = _model_snapshot(
tmp_path,