trying to increase accuracy and determinism

run_test.py

@@ -1,19 +1,30 @@
 #!/usr/bin/env python3
 import os, sys
-import accelerate, safetensors.torch, transformers, torch, tqdm
 
 model_id, revision = sys.argv[1:]
 user, model = model_id.split('/')
-
 fn = f'{user}_{model}_{revision}.logits.safetensors'
 
+import torch, numpy as np, random
+torch.backends.cuda.matmul.allow_fp16_reduced_precision_reduction = False
+torch.backends.cuda.matmul.allow_bf16_reduced_precision_reduction = False
+torch.backends.cudnn.allow_tf32 = False
+torch.backends.cudnn.benchmark = False
+torch.use_deterministic_algorithms(True)
+os.environ['CUBLAS_WORKSPACE_CONFIG'] = ':4096:2'
+os.environ['CUDA_LAUNCH_BLOCKING'] = '1'
+torch.manual_seed(0)
+random.seed(0)
+np.random.seed(0)
+import accelerate, safetensors.torch, transformers, torch, tqdm
+
 config = transformers.AutoConfig.from_pretrained(model_id, revision=revision, trust_remote_code=True)
 tokenizer = transformers.AutoTokenizer.from_pretrained(model_id, revision=revision, trust_remote_code=True)
 Model = transformers.AutoModelForCausalLM
 if config.model_type == 'deepseek_v3':
     #Model = transformers.DeepseekV3ForCausalLM
     pass
-model = Model.from_pretrained(model_id, revision=revision, trust_remote_code=True, torch_dtype='auto', device_map='cpu')
+model = Model.from_pretrained(model_id, revision=revision, trust_remote_code=True, torch_dtype=torch.float64, use_safetensors=True)
 if config.model_type == 'deepseek_v3':
     model._supports_cache_class = False
 model = accelerate.cpu_offload(model, 'cuda:0', offload_buffers=True)