Delete model_diagnostics.py
Browse files- model_diagnostics.py +0 -44
model_diagnostics.py
DELETED
|
@@ -1,44 +0,0 @@
|
|
| 1 |
-
import re
|
| 2 |
-
|
| 3 |
-
class ModelDiagnostics:
|
| 4 |
-
@staticmethod
|
| 5 |
-
def estimate_vram(param_str):
|
| 6 |
-
"""
|
| 7 |
-
Estimates VRAM usage based on parameter string (e.g., '7B', '0.5B').
|
| 8 |
-
Formula: (Params * Precision Bytes) + 20% Overhead for Context/Activations
|
| 9 |
-
"""
|
| 10 |
-
try:
|
| 11 |
-
# Clean string and extract number
|
| 12 |
-
clean_str = param_str.lower().replace('b', '').replace('m', '')
|
| 13 |
-
val = float(clean_str)
|
| 14 |
-
|
| 15 |
-
# Normalize to Billions
|
| 16 |
-
if 'm' in param_str.lower():
|
| 17 |
-
val = val / 1000.0
|
| 18 |
-
|
| 19 |
-
# Constants
|
| 20 |
-
overhead = 1.2 # 20% overhead for context window/activations
|
| 21 |
-
|
| 22 |
-
# Calculations
|
| 23 |
-
fp16_gb = (val * 2 * overhead) # 2 bytes per param
|
| 24 |
-
int8_gb = (val * 1 * overhead) # 1 byte per param
|
| 25 |
-
fp32_gb = (val * 4 * overhead) # 4 bytes per param
|
| 26 |
-
|
| 27 |
-
return {
|
| 28 |
-
"FP32 (Training/Full)": f"{fp32_gb:.2f} GB",
|
| 29 |
-
"FP16 (Inference)": f"{fp16_gb:.2f} GB",
|
| 30 |
-
"INT8 (Quantized)": f"{int8_gb:.2f} GB",
|
| 31 |
-
"params_in_billions": val
|
| 32 |
-
}
|
| 33 |
-
except Exception as e:
|
| 34 |
-
return None
|
| 35 |
-
|
| 36 |
-
@staticmethod
|
| 37 |
-
def get_layer_structure(model):
|
| 38 |
-
"""
|
| 39 |
-
Returns the raw string representation of the PyTorch model modules.
|
| 40 |
-
"""
|
| 41 |
-
if model:
|
| 42 |
-
# We strip the outer wrapper to get straight to the layers
|
| 43 |
-
return str(model)
|
| 44 |
-
return "Model not loaded."
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|