reduce memory footprint bfloat16
Browse files
app.py
CHANGED
@@ -19,8 +19,9 @@ model_id = "Tonic/c4ai-command-a-03-2025-4bit_fp4"
|
|
19 |
quantization_config = BitsAndBytesConfig(
|
20 |
load_in_4bit=True, # Enable 4-bit quantization
|
21 |
bnb_4bit_quant_type="fp4", # Use FP4 quantization
|
22 |
-
bnb_4bit_use_double_quant=True
|
23 |
# llm_int8_enable_fp32_cpu_offload=True # Allow CPU offloading for 32-bit modules
|
|
|
24 |
)
|
25 |
|
26 |
# Load tokenizer and model
|
@@ -31,7 +32,7 @@ model = AutoModelForCausalLM.from_pretrained(
|
|
31 |
# device_map="auto", # Automatically map to available devices
|
32 |
torch_dtype=torch.bfloat16,
|
33 |
token=HF_TOKEN,
|
34 |
-
max_position_embeddings=
|
35 |
)
|
36 |
|
37 |
@spaces.GPU
|
|
|
19 |
quantization_config = BitsAndBytesConfig(
|
20 |
load_in_4bit=True, # Enable 4-bit quantization
|
21 |
bnb_4bit_quant_type="fp4", # Use FP4 quantization
|
22 |
+
bnb_4bit_use_double_quant=True, # Optional: double quantization for better precision
|
23 |
# llm_int8_enable_fp32_cpu_offload=True # Allow CPU offloading for 32-bit modules
|
24 |
+
bnb_4bit_compute_dtype=torch.bfloat16 # Use bfloat16 for computation to save memory
|
25 |
)
|
26 |
|
27 |
# Load tokenizer and model
|
|
|
32 |
# device_map="auto", # Automatically map to available devices
|
33 |
torch_dtype=torch.bfloat16,
|
34 |
token=HF_TOKEN,
|
35 |
+
max_position_embeddings=4096 # Reduce context window to 8k tokens (from 128k)
|
36 |
)
|
37 |
|
38 |
@spaces.GPU
|