pytorch-rocm-llms: unify transformers pin and disable mmap for halo gpt-oss-20b

playbooks/core/pytorch-rocm-llms/README.md

@@ -110,15 +110,15 @@ pytorch-env\Scripts\activate
 <!-- @os:windows -->
 <!-- @test:id=install-deps timeout=300 setup=activate-venv -->
 ```bash
-pip install transformers==4.57.1 safetensors==0.6.2 accelerate sentencepiece protobuf
+pip install transformers==5.10.1 safetensors accelerate sentencepiece protobuf
 ```
 <!-- @test:end -->
 <!-- @os:end -->
 
 <!-- @os:linux -->
 <!-- @test:id=install-deps timeout=300 setup=activate-venv -->
 ```bash
-pip install transformers safetensors accelerate sentencepiece protobuf
+pip install transformers==5.10.1 safetensors accelerate sentencepiece protobuf
 ```
 <!-- @test:end -->
 <!-- @os:end -->
@@ -136,7 +136,7 @@ pip install transformers==5.10.1 safetensors accelerate sentencepiece protobuf
 <!-- @os:linux -->
 <!-- @test:id=install-deps timeout=300 setup=activate-venv -->
 ```bash
-pip install "transformers>=5.9.0" safetensors accelerate sentencepiece protobuf
+pip install transformers==5.10.1 safetensors accelerate sentencepiece protobuf
 ```
 <!-- @test:end -->
 <!-- @os:end -->
@@ -208,7 +208,8 @@ tokenizer = AutoTokenizer.from_pretrained(model_name)
 model = AutoModelForCausalLM.from_pretrained(
     model_name,
     torch_dtype=torch.bfloat16,
-    device_map="auto"
+    device_map="auto",
+    disable_mmap=True
 )
 ```
 <!-- @test:end -->

playbooks/core/pytorch-rocm-llms/assets/run_llm.py

@@ -48,6 +48,7 @@ def load_model(model_name):
             model_name,
             torch_dtype=torch.bfloat16,
             device_map="auto",
+            disable_mmap=True,
         )
     except (ValueError, KeyError):
         # Vision-language checkpoints (e.g. Qwen3.5) are not registered under
@@ -57,6 +58,7 @@ def load_model(model_name):
             model_name,
             torch_dtype=torch.bfloat16,
             device_map="auto",
+            disable_mmap=True,
         )
     return tokenizer, model
 

playbooks/core/pytorch-rocm-llms/assets/summarizer.py

@@ -65,6 +65,7 @@ def __init__(self, model=DEFAULT_MODEL):
                 self.model_name,
                 torch_dtype=torch.bfloat16,
                 device_map=device,  # whole model on one device; no CPU/meta offload
+                disable_mmap=True,
             )
         except (ValueError, KeyError):
             # Vision-language checkpoints (e.g. Qwen3.5) are not registered
@@ -74,6 +75,7 @@ def __init__(self, model=DEFAULT_MODEL):
                 self.model_name,
                 torch_dtype=torch.bfloat16,
                 device_map=device,  # whole model on one device; no CPU/meta offload
+                disable_mmap=True,
             )
         print("[OK] Model ready!\n")