oobabooga's picture
Update app.py
2c659e0 verified
raw
history blame
15.2 kB
import io
import re
import struct
from enum import IntEnum
from math import floor
import requests
import gradio as gr
class GGUFValueType(IntEnum):
UINT8 = 0
INT8 = 1
UINT16 = 2
INT16 = 3
UINT32 = 4
INT32 = 5
FLOAT32 = 6
BOOL = 7
STRING = 8
ARRAY = 9
UINT64 = 10
INT64 = 11
FLOAT64 = 12
_simple_value_packing = {
GGUFValueType.UINT8: "<B",
GGUFValueType.INT8: "<b",
GGUFValueType.UINT16: "<H",
GGUFValueType.INT16: "<h",
GGUFValueType.UINT32: "<I",
GGUFValueType.INT32: "<i",
GGUFValueType.FLOAT32: "<f",
GGUFValueType.UINT64: "<Q",
GGUFValueType.INT64: "<q",
GGUFValueType.FLOAT64: "<d",
GGUFValueType.BOOL: "?",
}
value_type_info = {
GGUFValueType.UINT8: 1,
GGUFValueType.INT8: 1,
GGUFValueType.UINT16: 2,
GGUFValueType.INT16: 2,
GGUFValueType.UINT32: 4,
GGUFValueType.INT32: 4,
GGUFValueType.FLOAT32: 4,
GGUFValueType.UINT64: 8,
GGUFValueType.INT64: 8,
GGUFValueType.FLOAT64: 8,
GGUFValueType.BOOL: 1,
}
def get_single(value_type, file):
if value_type == GGUFValueType.STRING:
value_length = struct.unpack("<Q", file.read(8))[0]
value = file.read(value_length)
try:
value = value.decode('utf-8')
except:
pass
else:
type_str = _simple_value_packing.get(value_type)
bytes_length = value_type_info.get(value_type)
value = struct.unpack(type_str, file.read(bytes_length))[0]
return value
def load_metadata_from_file(file_obj):
"""Load metadata from a file-like object"""
metadata = {}
GGUF_MAGIC = struct.unpack("<I", file_obj.read(4))[0]
GGUF_VERSION = struct.unpack("<I", file_obj.read(4))[0]
ti_data_count = struct.unpack("<Q", file_obj.read(8))[0]
kv_data_count = struct.unpack("<Q", file_obj.read(8))[0]
if GGUF_VERSION == 1:
raise Exception('You are using an outdated GGUF, please download a new one.')
for i in range(kv_data_count):
key_length = struct.unpack("<Q", file_obj.read(8))[0]
key = file_obj.read(key_length)
value_type = GGUFValueType(struct.unpack("<I", file_obj.read(4))[0])
if value_type == GGUFValueType.ARRAY:
ltype = GGUFValueType(struct.unpack("<I", file_obj.read(4))[0])
length = struct.unpack("<Q", file_obj.read(8))[0]
arr = [get_single(ltype, file_obj) for _ in range(length)]
metadata[key.decode()] = arr
else:
value = get_single(value_type, file_obj)
metadata[key.decode()] = value
# Extract specific fields needed for VRAM calculation
extracted_fields = {}
for key, value in metadata.items():
if key.endswith('.block_count'):
extracted_fields['n_layers'] = value
elif key.endswith('.attention.head_count_kv'):
extracted_fields['n_kv_heads'] = max(value) if isinstance(value, list) else value
elif key.endswith('.embedding_length'):
extracted_fields['embedding_dim'] = value
elif key.endswith('.context_length'):
extracted_fields['context_length'] = value
elif key.endswith('.feed_forward_length'):
extracted_fields['feed_forward_dim'] = value
# Add extracted fields to metadata for easy access
metadata.update(extracted_fields)
return metadata
def download_gguf_partial(url, max_bytes=25 * 1024 * 1024):
"""Download the first max_bytes from a GGUF URL"""
try:
# Set up headers for partial content request
headers = {'Range': f'bytes=0-{max_bytes-1}'}
# Make the request
response = requests.get(url, headers=headers, stream=True)
response.raise_for_status()
# Read the content
content = response.content
# Convert to BytesIO for file-like interface
return io.BytesIO(content)
except Exception as e:
raise Exception(f"Failed to download GGUF file: {str(e)}")
def load_metadata(model_url, current_metadata):
"""Load metadata from model URL and return updated metadata dict"""
if not model_url or model_url.strip() == "":
return {}, "Please enter a model URL"
try:
# Get model size first
model_size_mb = get_model_size_mb_from_url(model_url)
# Normalize URL for downloading
normalized_url = normalize_huggingface_url(model_url)
# Download the first 25MB of the file
file_obj = download_gguf_partial(normalized_url)
# Parse the metadata
metadata = load_metadata_from_file(file_obj)
# Extract filename from URL
gguf_filename = model_url.split('/')[-1].split('?')[0] # Remove query parameters if any
# Extract model name from URL if it's a Hugging Face URL
model_name = model_url
if "huggingface.co/" in model_url:
try:
# Extract model name from URL like https://huggingface.co/user/model
parts = model_url.split("huggingface.co/")[1].split("/")
if len(parts) >= 2:
model_name = f"{parts[0]}/{parts[1]}"
except:
model_name = model_url
# Add URL, model name, and size to metadata
metadata['url'] = model_url
metadata['model_name'] = model_name
metadata['model_size_mb'] = model_size_mb
metadata['loaded'] = True
return metadata, gr.update(value=metadata["n_layers"], maximum=metadata["n_layers"]), f"Metadata loaded successfully for: {gguf_filename}"
except Exception as e:
error_msg = f"Error loading metadata: {str(e)}"
return {}, gr.update(), error_msg
def normalize_huggingface_url(url: str) -> str:
"""Normalize HuggingFace URL to resolve format for direct access"""
if 'huggingface.co' not in url:
return url
# Remove query parameters first
base_url = url.split('?')[0]
# Convert blob URL to resolve URL
if '/blob/' in base_url:
base_url = base_url.replace('/blob/', '/resolve/')
return base_url
def get_model_size_mb_from_url(model_url: str) -> float:
"""Get model size in MB from URL without downloading, handling multi-part files"""
try:
# Normalize the URL for direct access
normalized_url = normalize_huggingface_url(model_url)
# Get size of the main file
response = requests.head(normalized_url, allow_redirects=True)
response.raise_for_status()
main_file_size = int(response.headers.get('content-length', 0))
# Extract filename from original URL
filename = normalized_url.split('/')[-1]
# Check for multipart pattern (e.g., model-00001-of-00002.gguf)
match = re.match(r'(.+)-(\d+)-of-(\d+)\.gguf$', filename)
if match:
base_pattern = match.group(1)
total_parts = int(match.group(3))
total_size = 0
base_url = '/'.join(normalized_url.split('/')[:-1]) + '/'
# Get size of all parts
for part_num in range(1, total_parts + 1):
part_filename = f"{base_pattern}-{part_num:05d}-of-{total_parts:05d}.gguf"
part_url = base_url + part_filename
try:
part_response = requests.head(part_url, allow_redirects=True)
part_response.raise_for_status()
part_size = int(part_response.headers.get('content-length', 0))
total_size += part_size
except requests.RequestException as e:
print(f"Warning: Could not get size of {part_filename}, estimating...")
# If we can't get some parts, estimate based on what we have
if total_size > 0:
avg_size = total_size / (part_num - 1)
remaining_parts = total_parts - (part_num - 1)
total_size += avg_size * remaining_parts
else:
# Fallback to main file size * total parts
total_size = main_file_size * total_parts
break
return total_size / (1024 ** 2)
else:
# Single part file
return main_file_size / (1024 ** 2)
except Exception as e:
print(f"Error getting model size: {e}")
return 0.0
def estimate_vram(metadata, gpu_layers, ctx_size, cache_type):
"""Calculate VRAM usage using the actual formula"""
try:
# Extract required values from metadata
n_layers = metadata.get('n_layers')
n_kv_heads = metadata.get('n_kv_heads')
embedding_dim = metadata.get('embedding_dim')
context_length = metadata.get('context_length')
feed_forward_dim = metadata.get('feed_forward_dim')
size_in_mb = metadata.get('model_size_mb', 0)
# Check if we have all required fields
required_fields = [n_layers, n_kv_heads, embedding_dim, context_length, feed_forward_dim]
if any(field is None for field in required_fields):
missing = [name for name, field in zip(
['n_layers', 'n_kv_heads', 'embedding_dim', 'context_length', 'feed_forward_dim'],
required_fields) if field is None]
raise ValueError(f"Missing required metadata fields: {missing}")
# Ensure gpu_layers doesn't exceed total layers
if gpu_layers > n_layers:
gpu_layers = n_layers
# Convert cache_type to numeric
if cache_type == 'q4_0':
cache_type = 4
elif cache_type == 'q8_0':
cache_type = 8
else:
cache_type = 16
# Derived features
size_per_layer = size_in_mb / max(n_layers, 1e-6)
kv_cache_factor = n_kv_heads * cache_type * ctx_size
embedding_per_context = embedding_dim / ctx_size
# Calculate VRAM using the model
# Details: https://oobabooga.github.io/blog/posts/gguf-vram-formula/
vram = (
(size_per_layer - 17.99552795246051 + 3.148552680382576e-05 * kv_cache_factor)
* (gpu_layers + max(0.9690636483914102, cache_type - (floor(50.77817218646521 * embedding_per_context) + 9.987899908205632)))
+ 1516.522943869404
)
return vram
except Exception as e:
print(f"Error in VRAM calculation: {e}")
raise
def estimate_vram_wrapper(model_metadata, gpu_layers, ctx_size, cache_type):
"""Wrapper function to estimate VRAM usage"""
if not model_metadata or 'model_name' not in model_metadata:
return "<div id=\"vram-info\">Estimated VRAM to load the model:</div>"
# Use cache_type directly (it's already a string from the radio button)
try:
result = estimate_vram(model_metadata, gpu_layers, ctx_size, cache_type)
conservative = result + 577
return f"""<div id="vram-info">
<div>Expected VRAM usage: <span class="value">{result:.0f} MiB</span></div>
<div>Safe estimate: <span class="value">{conservative:.0f} MiB</span> - 95% chance the VRAM is at most this.</div>
</div>"""
except Exception as e:
return f"<div id=\"vram-info\">Estimated VRAM to load the model: <span class=\"value\">Error: {str(e)}</span></div>"
def create_ui():
"""Create the simplified UI"""
# Custom CSS to limit max width and center the content
css = """
body {
max-width: 810px !important;
margin: 0 auto !important;
}
#vram-info {
padding: 10px;
border-radius: 4px;
background-color: var(--background-fill-secondary);
}
#vram-info .value {
font-weight: bold;
color: var(--primary-500);
}
"""
with gr.Blocks(css=css) as demo:
# State to hold model metadata
model_metadata = gr.State(value={})
gr.Markdown("# Accurate GGUF VRAM Calculator\n\nCalculate VRAM for GGUF models from GPU layers and context length using an accurate formula.\n\nFor an explanation about how this works, consult this blog post: https://oobabooga.github.io/blog/posts/gguf-vram-formula/")
with gr.Row():
with gr.Column():
# Model URL input
model_url = gr.Textbox(
label="GGUF Model URL",
placeholder="https://huggingface.co/unsloth/Qwen3-235B-A22B-GGUF/blob/main/UD-Q2_K_XL/Qwen3-235B-A22B-UD-Q2_K_XL-00001-of-00002.gguf",
value=""
)
# Load metadata button
load_metadata_btn = gr.Button("Load metadata", elem_classes='refresh-button')
# GPU layers slider
gpu_layers = gr.Slider(
label="GPU Layers",
minimum=0,
maximum=256,
value=256,
info='`--gpu-layers` in llama.cpp.'
)
# Context size slider
ctx_size = gr.Slider(
label='Context Length',
minimum=512,
maximum=131072,
step=256,
value=8192,
info='`--ctx-size` in llama.cpp.'
)
# Cache type checkbox group
cache_type = gr.Radio(
choices=['fp16', 'q8_0', 'q4_0'],
value='fp16',
label="Cache Type",
info='Cache quantization.'
)
# VRAM info display
vram_info = gr.HTML(
value="<div id=\"vram-info\">Estimated VRAM to load the model:</div>"
)
# Status display
status = gr.Textbox(
label="Status",
value="No model loaded",
interactive=False
)
# Event handlers
load_metadata_btn.click(
load_metadata,
inputs=[model_url, model_metadata],
outputs=[model_metadata, gpu_layers, status],
show_progress=True
).then(
estimate_vram_wrapper,
inputs=[model_metadata, gpu_layers, ctx_size, cache_type],
outputs=[vram_info],
show_progress=False
)
# Update VRAM estimate when any parameter changes
for component in [gpu_layers, ctx_size, cache_type]:
component.change(
estimate_vram_wrapper,
inputs=[model_metadata, gpu_layers, ctx_size, cache_type],
outputs=[vram_info],
show_progress=False
)
# Also update when model_metadata state changes
model_metadata.change(
estimate_vram_wrapper,
inputs=[model_metadata, gpu_layers, ctx_size, cache_type],
outputs=[vram_info],
show_progress=False
)
return demo
if __name__ == "__main__":
# Create and launch the app
demo = create_ui()
demo.launch()