notes from deepseek attempt

run_test.py

@@ -5,9 +5,13 @@ STORE_WEIGHTS = False
 DEDUPLICATE_SAFETENSORS = True
 SAVE_ON_CRASH = False
 FAKE_H100 = False
+# notes transformers 4.49.0:
+# - a model with unexpected weights must not have 'disk' anywhere in device_map
+# - a model with dtypes unsupported by numpy must have offload_state_dict = False
+# - dtype=float64 can overflow, no dtype specified at transformers/modeling_attn_mask_utils.py(158)_make_causal_mask() defaulted to float32
 TORCH_DTYPE = 'float64'
 USE_GPU = False
-DEVICE_MAP = 'auto' # 'cuda' for deepseek
+DEVICE_MAP = 'auto'
 model_id, revision = sys.argv[1:]
 user, model = model_id.split('/')
 prompt = 'Once upon a time,'
@@ -64,7 +68,7 @@ if config.model_type == 'deepseek_v3':
 max_memory = accelerate.utils.get_max_memory()
 if not USE_GPU:
     max_memory = {'cpu': max_memory['cpu']}
-max_memory['cpu'] //= 3
+#max_memory['cpu'] //= 3
 model_kwparams = dict(pretrained_model_name_or_path=model_id, revision=revision, trust_remote_code=True,
     torch_dtype=getattr(torch, TORCH_DTYPE, TORCH_DTYPE), use_safetensors=True, low_cpu_mem_usage=True,
     device_map=DEVICE_MAP, max_memory=max_memory, offload_state_dict=True, offload_buffers=True)