Merge branch 'main' into nightly

pyproject.toml

@@ -37,7 +37,7 @@ triton = [
 ]
 
 huggingface = [
-    "unsloth_zoo>=2025.5.1",
+    "unsloth_zoo>=2025.5.5",
     "packaging",
     "tyro",
     "transformers==4.51.3,!=4.47.0",
@@ -381,7 +381,7 @@ colab-ampere-torch220 = [
     "flash-attn>=2.6.3",
 ]
 colab-new = [
-    "unsloth_zoo>=2025.5.1",
+    "unsloth_zoo>=2025.5.5",
     "packaging",
     "tyro",
     "transformers==4.51.3,!=4.47.0",

unsloth/dataprep/synthetic.py

@@ -64,7 +64,7 @@ class SyntheticDataKit:
             model_name,
             token = token,
         )
-        patch_vllm()
+        patch_vllm(debug = False)
         engine_args = load_vllm(
             model_name             = model_name,
             config                 = self.config,
@@ -81,6 +81,9 @@ class SyntheticDataKit:
 
         if "device" in engine_args: del engine_args["device"]
         if "model"  in engine_args: del engine_args["model"]
+        if "compilation_config" in engine_args:
+            # Cannot parse in vllm serve
+            engine_args["compilation_config"] = 3
 
         subprocess_commands = [
             "vllm", "serve", str(model_name),
@@ -113,7 +116,8 @@ class SyntheticDataKit:
                 print("Stdout stream ended before readiness message detected.")
                 break
             output_str = output.decode('utf-8', errors='ignore').strip()
-            print(f"vLLM STDOUT: {output_str}")
+            if "platform is" not in output_str:
+                print(f"vLLM STDOUT: {output_str}")
             if ready_message_part in output:
                 print(f"\n--- vLLM Server Ready (Detected: '{ready_message_part.decode()}') ---")
                 ready = True
@@ -192,12 +196,7 @@ class SyntheticDataKit:
             gc.collect()
 
         # Delete vLLM module as well
-        # We delete llm.llm_engine.model_executor, so first make it accessible
-        class Dummy0: model_executor = 1
-        class Dummy1: llm_engine = Dummy0()
-        class Dummy2: llm = Dummy1()
-        llm = Dummy2().llm.llm_engine.model_executor
-        delete_vllm(llm)
+        delete_vllm(llm = None)
     pass
 
     def __enter__(self): return self

unsloth/models/_utils.py

@@ -12,7 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-__version__ = "2025.5.1"
+__version__ = "2025.5.3"
 
 __all__ = [
     "SUPPORTS_BFLOAT16",

unsloth/models/loader.py

@@ -540,6 +540,9 @@ class FastModel(FastBaseModel):
             os.environ["UNSLOTH_COMPILE_DISABLE"] = "1"
             if transformers_version < Version("4.50.0.dev0"):
                 raise RuntimeError("Unsloth: Granite Vision only works on transformers >= 4.50.0." + NIGHTLY)
+        elif "csm-1b" in model_name.lower():
+            os.environ["UNSLOTH_COMPILE_DISABLE"] = "1"
+            os.environ["UNSLOTH_DISABLE_FAST_GENERATION"] = "1"
         elif "olmo-2" in model_name.lower() and transformers_version < Version("4.50.0.dev0"):
             raise RuntimeError("Unsloth: OLMo-2 only works on transformers >= 4.50.0." + NIGHTLY)
         pass