jkottu's picture
Claude Opus 4.5
Initial commit: LLM Inference Dashboard
aefabf0
Raw History Blame Contribute Delete
3.04 kB
"""Quantization details panel component."""
import gradio as gr
import pandas as pd
from typing import Dict, Any, Tuple, Optional
from collectors.quant_collector import QuantizationCollector, QuantizationInfo
def create_quant_panel() -> Dict[str, Any]:
"""
Create the quantization details panel.
Returns:
Dictionary of Gradio components
"""
with gr.Column():
gr.Markdown("### Quantization Details")
with gr.Row():
quant_type = gr.Textbox(
label="Quantization Method",
interactive=False,
)
bits = gr.Number(
label="Bits",
precision=0,
interactive=False,
)
group_size = gr.Number(
label="Group Size",
precision=0,
interactive=False,
)
# Full configuration JSON
quant_details = gr.JSON(
label="Full Configuration",
)
# Layer precision table
gr.Markdown("#### Per-Layer Precision")
layer_table = gr.Dataframe(
headers=["Layer", "Bits", "Group Size", "Dtype"],
datatype=["str", "number", "str", "str"],
label="Layer Precisions",
interactive=False,
)
return {
"quant_type": quant_type,
"bits": bits,
"group_size": group_size,
"quant_details": quant_details,
"layer_table": layer_table,
}
def update_quant_panel(
collector: QuantizationCollector,
) -> Tuple[str, int, Optional[int], Dict, pd.DataFrame]:
"""
Update the quantization panel with current data.
Args:
collector: Quantization collector instance
Returns:
Tuple of (method, bits, group_size, details_json, layer_table)
"""
info = collector.detect()
layers = collector.get_layer_precisions()
# Build layer table
layer_rows = []
for layer in layers[:20]: # Limit to 20 rows
layer_rows.append({
"Layer": layer.layer_name,
"Bits": layer.bits,
"Group Size": str(layer.group_size) if layer.group_size else "-",
"Dtype": layer.dtype,
})
layer_df = pd.DataFrame(layer_rows) if layer_rows else pd.DataFrame(
columns=["Layer", "Bits", "Group Size", "Dtype"]
)
return (
info.method,
info.bits,
info.group_size,
info.to_dict(),
layer_df,
)
def get_quant_summary(info: QuantizationInfo) -> str:
"""
Get a summary string for the quantization.
Args:
info: QuantizationInfo instance
Returns:
Human-readable summary string
"""
if info.method == "None (FP16/BF16)":
return f"Full precision ({info.compute_dtype or 'float16'})"
summary = f"{info.method} {info.bits}-bit"
if info.group_size:
summary += f", group size {info.group_size}"
if info.quant_type:
summary += f" ({info.quant_type})"
return summary