Spaces:
Sleeping
Sleeping
Commit
·
8049238
1
Parent(s):
be6c085
memory management
Browse files- app.py +20 -2
- cognitive_mapping_probe/auto_experiment.py +13 -22
- cognitive_mapping_probe/orchestrator_seismograph.py +21 -19
app.py
CHANGED
|
@@ -1,6 +1,8 @@
|
|
| 1 |
import gradio as gr
|
| 2 |
import pandas as pd
|
| 3 |
import traceback
|
|
|
|
|
|
|
| 4 |
|
| 5 |
from cognitive_mapping_probe.orchestrator_seismograph import run_seismic_analysis
|
| 6 |
from cognitive_mapping_probe.auto_experiment import run_auto_suite, get_curated_experiments
|
|
@@ -9,6 +11,14 @@ from cognitive_mapping_probe.utils import dbg
|
|
| 9 |
|
| 10 |
theme = gr.themes.Soft(primary_hue="indigo", secondary_hue="blue").set(body_background_fill="#f0f4f9", block_background_fill="white")
|
| 11 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 12 |
def run_single_analysis_display(*args, progress=gr.Progress(track_tqdm=True)):
|
| 13 |
"""Wrapper für ein einzelnes manuelles Experiment."""
|
| 14 |
try:
|
|
@@ -17,16 +27,26 @@ def run_single_analysis_display(*args, progress=gr.Progress(track_tqdm=True)):
|
|
| 17 |
deltas = results.get("state_deltas", [])
|
| 18 |
df = pd.DataFrame({"Internal Step": range(len(deltas)), "State Change (Delta)": deltas})
|
| 19 |
stats_md = f"### Statistical Signature\n- **Mean Delta:** {stats.get('mean_delta', 0):.4f}\n- **Std Dev Delta:** {stats.get('std_delta', 0):.4f}\n- **Max Delta:** {stats.get('max_delta', 0):.4f}\n"
|
|
|
|
|
|
|
|
|
|
|
|
|
| 20 |
return f"{results.get('verdict', 'Error')}\n\n{stats_md}", df, results
|
| 21 |
except Exception:
|
|
|
|
| 22 |
return f"### ❌ Analysis Failed\n```\n{traceback.format_exc()}\n```", pd.DataFrame(), {}
|
| 23 |
|
| 24 |
def run_auto_suite_display(model_id, num_steps, seed, experiment_name, progress=gr.Progress(track_tqdm=True)):
|
| 25 |
"""Wrapper für die automatisierte Experiment-Suite mit Visualisierung."""
|
| 26 |
try:
|
| 27 |
summary_df, plot_df, all_results = run_auto_suite(model_id, int(num_steps), int(seed), experiment_name, progress)
|
|
|
|
|
|
|
|
|
|
|
|
|
| 28 |
return summary_df, plot_df, all_results
|
| 29 |
except Exception:
|
|
|
|
| 30 |
return pd.DataFrame(), pd.DataFrame(), f"### ❌ Auto-Experiment Failed\n```\n{traceback.format_exc()}\n```"
|
| 31 |
|
| 32 |
with gr.Blocks(theme=theme, title="Cognitive Seismograph 2.2") as demo:
|
|
@@ -49,7 +69,6 @@ with gr.Blocks(theme=theme, title="Cognitive Seismograph 2.2") as demo:
|
|
| 49 |
with gr.Column(scale=2):
|
| 50 |
gr.Markdown("### Single Run Results")
|
| 51 |
manual_verdict = gr.Markdown("Die Analyse erscheint hier.")
|
| 52 |
-
# KORREKTUR: `interactive=True` für Legende hinzugefügt
|
| 53 |
manual_plot = gr.LinePlot(x="Internal Step", y="State Change (Delta)", title="Internal State Dynamics", show_label=True, height=400, interactive=True)
|
| 54 |
with gr.Accordion("Raw JSON Output", open=False):
|
| 55 |
manual_raw_json = gr.JSON()
|
|
@@ -72,7 +91,6 @@ with gr.Blocks(theme=theme, title="Cognitive Seismograph 2.2") as demo:
|
|
| 72 |
auto_run_btn = gr.Button("Run Curated Auto-Experiment", variant="primary")
|
| 73 |
with gr.Column(scale=2):
|
| 74 |
gr.Markdown("### Suite Results Summary")
|
| 75 |
-
# KORREKTUR: `interactive=True` für Legende hinzugefügt
|
| 76 |
auto_plot_output = gr.LinePlot(
|
| 77 |
x="Step", y="Delta", color="Experiment",
|
| 78 |
title="Comparative Cognitive Dynamics",
|
|
|
|
| 1 |
import gradio as gr
|
| 2 |
import pandas as pd
|
| 3 |
import traceback
|
| 4 |
+
import gc
|
| 5 |
+
import torch
|
| 6 |
|
| 7 |
from cognitive_mapping_probe.orchestrator_seismograph import run_seismic_analysis
|
| 8 |
from cognitive_mapping_probe.auto_experiment import run_auto_suite, get_curated_experiments
|
|
|
|
| 11 |
|
| 12 |
theme = gr.themes.Soft(primary_hue="indigo", secondary_hue="blue").set(body_background_fill="#f0f4f9", block_background_fill="white")
|
| 13 |
|
| 14 |
+
def cleanup_memory():
|
| 15 |
+
"""Eine zentrale Funktion zum Aufräumen des Speichers nach einem Lauf."""
|
| 16 |
+
dbg("Cleaning up memory...")
|
| 17 |
+
gc.collect()
|
| 18 |
+
if torch.cuda.is_available():
|
| 19 |
+
torch.cuda.empty_cache()
|
| 20 |
+
dbg("Memory cleanup complete.")
|
| 21 |
+
|
| 22 |
def run_single_analysis_display(*args, progress=gr.Progress(track_tqdm=True)):
|
| 23 |
"""Wrapper für ein einzelnes manuelles Experiment."""
|
| 24 |
try:
|
|
|
|
| 27 |
deltas = results.get("state_deltas", [])
|
| 28 |
df = pd.DataFrame({"Internal Step": range(len(deltas)), "State Change (Delta)": deltas})
|
| 29 |
stats_md = f"### Statistical Signature\n- **Mean Delta:** {stats.get('mean_delta', 0):.4f}\n- **Std Dev Delta:** {stats.get('std_delta', 0):.4f}\n- **Max Delta:** {stats.get('max_delta', 0):.4f}\n"
|
| 30 |
+
|
| 31 |
+
# WICHTIG: Speicher aufräumen, BEVOR die Ergebnisse an Gradio zurückgegeben werden.
|
| 32 |
+
cleanup_memory()
|
| 33 |
+
|
| 34 |
return f"{results.get('verdict', 'Error')}\n\n{stats_md}", df, results
|
| 35 |
except Exception:
|
| 36 |
+
cleanup_memory()
|
| 37 |
return f"### ❌ Analysis Failed\n```\n{traceback.format_exc()}\n```", pd.DataFrame(), {}
|
| 38 |
|
| 39 |
def run_auto_suite_display(model_id, num_steps, seed, experiment_name, progress=gr.Progress(track_tqdm=True)):
|
| 40 |
"""Wrapper für die automatisierte Experiment-Suite mit Visualisierung."""
|
| 41 |
try:
|
| 42 |
summary_df, plot_df, all_results = run_auto_suite(model_id, int(num_steps), int(seed), experiment_name, progress)
|
| 43 |
+
|
| 44 |
+
# WICHTIG: Speicher auch hier aufräumen.
|
| 45 |
+
cleanup_memory()
|
| 46 |
+
|
| 47 |
return summary_df, plot_df, all_results
|
| 48 |
except Exception:
|
| 49 |
+
cleanup_memory()
|
| 50 |
return pd.DataFrame(), pd.DataFrame(), f"### ❌ Auto-Experiment Failed\n```\n{traceback.format_exc()}\n```"
|
| 51 |
|
| 52 |
with gr.Blocks(theme=theme, title="Cognitive Seismograph 2.2") as demo:
|
|
|
|
| 69 |
with gr.Column(scale=2):
|
| 70 |
gr.Markdown("### Single Run Results")
|
| 71 |
manual_verdict = gr.Markdown("Die Analyse erscheint hier.")
|
|
|
|
| 72 |
manual_plot = gr.LinePlot(x="Internal Step", y="State Change (Delta)", title="Internal State Dynamics", show_label=True, height=400, interactive=True)
|
| 73 |
with gr.Accordion("Raw JSON Output", open=False):
|
| 74 |
manual_raw_json = gr.JSON()
|
|
|
|
| 91 |
auto_run_btn = gr.Button("Run Curated Auto-Experiment", variant="primary")
|
| 92 |
with gr.Column(scale=2):
|
| 93 |
gr.Markdown("### Suite Results Summary")
|
|
|
|
| 94 |
auto_plot_output = gr.LinePlot(
|
| 95 |
x="Step", y="Delta", color="Experiment",
|
| 96 |
title="Comparative Cognitive Dynamics",
|
cognitive_mapping_probe/auto_experiment.py
CHANGED
|
@@ -1,4 +1,6 @@
|
|
| 1 |
import pandas as pd
|
|
|
|
|
|
|
| 2 |
from typing import Dict, List, Tuple
|
| 3 |
|
| 4 |
from .llm_iface import get_or_load_model
|
|
@@ -8,7 +10,6 @@ from .utils import dbg
|
|
| 8 |
def get_curated_experiments() -> Dict[str, List[Dict]]:
|
| 9 |
"""
|
| 10 |
Definiert die vordefinierten, wissenschaftlichen Experiment-Protokolle.
|
| 11 |
-
ERWEITERT um zusätzliche, aussagekräftige Experimente.
|
| 12 |
"""
|
| 13 |
experiments = {
|
| 14 |
"Calm vs. Chaos": [
|
|
@@ -61,6 +62,7 @@ def run_auto_suite(
|
|
| 61 |
summary_data = []
|
| 62 |
plot_data_frames = []
|
| 63 |
|
|
|
|
| 64 |
llm = get_or_load_model(model_id, seed)
|
| 65 |
|
| 66 |
total_runs = len(protocol)
|
|
@@ -69,40 +71,29 @@ def run_auto_suite(
|
|
| 69 |
dbg(f"--- Running Auto-Experiment: '{label}' ({i+1}/{total_runs}) ---")
|
| 70 |
|
| 71 |
results = run_seismic_analysis(
|
| 72 |
-
model_id=model_id,
|
| 73 |
-
|
| 74 |
-
|
| 75 |
-
num_steps=num_steps,
|
| 76 |
-
concept_to_inject=run_spec["concept"],
|
| 77 |
-
injection_strength=run_spec["strength"],
|
| 78 |
progress_callback=progress_callback,
|
| 79 |
-
llm_instance=llm
|
| 80 |
)
|
| 81 |
|
| 82 |
all_results[label] = results
|
| 83 |
stats = results.get("stats", {})
|
| 84 |
|
| 85 |
-
summary_data.append({
|
| 86 |
-
"Experiment": label,
|
| 87 |
-
"Mean Delta": stats.get("mean_delta"),
|
| 88 |
-
"Std Dev Delta": stats.get("std_delta"),
|
| 89 |
-
"Max Delta": stats.get("max_delta"),
|
| 90 |
-
})
|
| 91 |
|
| 92 |
deltas = results.get("state_deltas", [])
|
| 93 |
-
|
| 94 |
-
# Verwende das Label direkt für die Legende. Der vorherige Fix war nicht nötig,
|
| 95 |
-
# solange die Labels einzigartig pro Experiment sind.
|
| 96 |
-
df = pd.DataFrame({
|
| 97 |
-
"Step": range(len(deltas)),
|
| 98 |
-
"Delta": deltas,
|
| 99 |
-
"Experiment": label
|
| 100 |
-
})
|
| 101 |
plot_data_frames.append(df)
|
| 102 |
|
| 103 |
summary_df = pd.DataFrame(summary_data)
|
| 104 |
plot_df = pd.concat(plot_data_frames, ignore_index=True) if plot_data_frames else pd.DataFrame()
|
| 105 |
|
|
|
|
| 106 |
del llm
|
|
|
|
|
|
|
|
|
|
| 107 |
|
| 108 |
return summary_df, plot_df, all_results
|
|
|
|
| 1 |
import pandas as pd
|
| 2 |
+
import torch
|
| 3 |
+
import gc
|
| 4 |
from typing import Dict, List, Tuple
|
| 5 |
|
| 6 |
from .llm_iface import get_or_load_model
|
|
|
|
| 10 |
def get_curated_experiments() -> Dict[str, List[Dict]]:
|
| 11 |
"""
|
| 12 |
Definiert die vordefinierten, wissenschaftlichen Experiment-Protokolle.
|
|
|
|
| 13 |
"""
|
| 14 |
experiments = {
|
| 15 |
"Calm vs. Chaos": [
|
|
|
|
| 62 |
summary_data = []
|
| 63 |
plot_data_frames = []
|
| 64 |
|
| 65 |
+
# Lade das Modell einmal zu Beginn der Suite
|
| 66 |
llm = get_or_load_model(model_id, seed)
|
| 67 |
|
| 68 |
total_runs = len(protocol)
|
|
|
|
| 71 |
dbg(f"--- Running Auto-Experiment: '{label}' ({i+1}/{total_runs}) ---")
|
| 72 |
|
| 73 |
results = run_seismic_analysis(
|
| 74 |
+
model_id=model_id, prompt_type=run_spec["prompt_type"],
|
| 75 |
+
seed=seed, num_steps=num_steps,
|
| 76 |
+
concept_to_inject=run_spec["concept"], injection_strength=run_spec["strength"],
|
|
|
|
|
|
|
|
|
|
| 77 |
progress_callback=progress_callback,
|
| 78 |
+
llm_instance=llm # Wiederverwende die geladene LLM-Instanz
|
| 79 |
)
|
| 80 |
|
| 81 |
all_results[label] = results
|
| 82 |
stats = results.get("stats", {})
|
| 83 |
|
| 84 |
+
summary_data.append({ "Experiment": label, "Mean Delta": stats.get("mean_delta"), "Std Dev Delta": stats.get("std_delta"), "Max Delta": stats.get("max_delta"), })
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 85 |
|
| 86 |
deltas = results.get("state_deltas", [])
|
| 87 |
+
df = pd.DataFrame({ "Step": range(len(deltas)), "Delta": deltas, "Experiment": f"{i}: {label}" })
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 88 |
plot_data_frames.append(df)
|
| 89 |
|
| 90 |
summary_df = pd.DataFrame(summary_data)
|
| 91 |
plot_df = pd.concat(plot_data_frames, ignore_index=True) if plot_data_frames else pd.DataFrame()
|
| 92 |
|
| 93 |
+
# WICHTIG: Explizites Aufräumen am Ende der gesamten Suite
|
| 94 |
del llm
|
| 95 |
+
gc.collect()
|
| 96 |
+
if torch.cuda.is_available():
|
| 97 |
+
torch.cuda.empty_cache()
|
| 98 |
|
| 99 |
return summary_df, plot_df, all_results
|
cognitive_mapping_probe/orchestrator_seismograph.py
CHANGED
|
@@ -1,5 +1,6 @@
|
|
| 1 |
import torch
|
| 2 |
import numpy as np
|
|
|
|
| 3 |
from typing import Dict, Any, Optional
|
| 4 |
|
| 5 |
from .llm_iface import get_or_load_model
|
|
@@ -15,39 +16,35 @@ def run_seismic_analysis(
|
|
| 15 |
concept_to_inject: str,
|
| 16 |
injection_strength: float,
|
| 17 |
progress_callback,
|
| 18 |
-
llm_instance: Optional[Any] = None
|
| 19 |
) -> Dict[str, Any]:
|
| 20 |
"""
|
| 21 |
-
Orchestriert eine einzelne seismische Analyse.
|
| 22 |
-
|
| 23 |
"""
|
| 24 |
-
|
| 25 |
if llm_instance is None:
|
| 26 |
progress_callback(0.1, desc="Loading model...")
|
| 27 |
llm = get_or_load_model(model_id, seed)
|
| 28 |
-
|
| 29 |
else:
|
| 30 |
llm = llm_instance
|
| 31 |
-
llm.set_all_seeds(seed)
|
| 32 |
-
created_llm = False
|
| 33 |
|
| 34 |
injection_vector = None
|
| 35 |
if concept_to_inject and concept_to_inject.strip():
|
| 36 |
-
if not
|
| 37 |
injection_vector = get_concept_vector(llm, concept_to_inject.strip())
|
| 38 |
|
| 39 |
-
if not
|
| 40 |
|
| 41 |
state_deltas = run_silent_cogitation_seismic(
|
| 42 |
-
llm=llm,
|
| 43 |
-
|
| 44 |
-
|
| 45 |
-
temperature=0.1,
|
| 46 |
-
injection_vector=injection_vector,
|
| 47 |
-
injection_strength=injection_strength
|
| 48 |
)
|
| 49 |
|
| 50 |
-
if not
|
| 51 |
|
| 52 |
if state_deltas:
|
| 53 |
deltas_np = np.array(state_deltas)
|
|
@@ -60,9 +57,14 @@ def run_seismic_analysis(
|
|
| 60 |
|
| 61 |
results = { "verdict": verdict, "stats": stats, "state_deltas": state_deltas }
|
| 62 |
|
| 63 |
-
# Gib das Modell nur frei, wenn es in dieser
|
| 64 |
-
|
|
|
|
|
|
|
| 65 |
del llm
|
| 66 |
-
|
|
|
|
|
|
|
|
|
|
| 67 |
|
| 68 |
return results
|
|
|
|
| 1 |
import torch
|
| 2 |
import numpy as np
|
| 3 |
+
import gc
|
| 4 |
from typing import Dict, Any, Optional
|
| 5 |
|
| 6 |
from .llm_iface import get_or_load_model
|
|
|
|
| 16 |
concept_to_inject: str,
|
| 17 |
injection_strength: float,
|
| 18 |
progress_callback,
|
| 19 |
+
llm_instance: Optional[Any] = None
|
| 20 |
) -> Dict[str, Any]:
|
| 21 |
"""
|
| 22 |
+
Orchestriert eine einzelne seismische Analyse. Stellt sicher, dass das Modell
|
| 23 |
+
nur dann entladen wird, wenn es auch hier geladen wurde.
|
| 24 |
"""
|
| 25 |
+
local_llm_instance = False
|
| 26 |
if llm_instance is None:
|
| 27 |
progress_callback(0.1, desc="Loading model...")
|
| 28 |
llm = get_or_load_model(model_id, seed)
|
| 29 |
+
local_llm_instance = True
|
| 30 |
else:
|
| 31 |
llm = llm_instance
|
| 32 |
+
llm.set_all_seeds(seed)
|
|
|
|
| 33 |
|
| 34 |
injection_vector = None
|
| 35 |
if concept_to_inject and concept_to_inject.strip():
|
| 36 |
+
if not local_llm_instance: progress_callback(0.2, desc=f"Vectorizing '{concept_to_inject}'...")
|
| 37 |
injection_vector = get_concept_vector(llm, concept_to_inject.strip())
|
| 38 |
|
| 39 |
+
if not local_llm_instance: progress_callback(0.3, desc=f"Recording dynamics...")
|
| 40 |
|
| 41 |
state_deltas = run_silent_cogitation_seismic(
|
| 42 |
+
llm=llm, prompt_type=prompt_type,
|
| 43 |
+
num_steps=num_steps, temperature=0.1,
|
| 44 |
+
injection_vector=injection_vector, injection_strength=injection_strength
|
|
|
|
|
|
|
|
|
|
| 45 |
)
|
| 46 |
|
| 47 |
+
if not local_llm_instance: progress_callback(0.9, desc="Analyzing...")
|
| 48 |
|
| 49 |
if state_deltas:
|
| 50 |
deltas_np = np.array(state_deltas)
|
|
|
|
| 57 |
|
| 58 |
results = { "verdict": verdict, "stats": stats, "state_deltas": state_deltas }
|
| 59 |
|
| 60 |
+
# WICHTIG: Gib das Modell und den Speicher nur frei, wenn es in dieser
|
| 61 |
+
# Funktion auch erstellt wurde. Ansonsten ist die übergeordnete Funktion
|
| 62 |
+
# (z.B. `run_auto_suite`) für das Speichermanagement verantwortlich.
|
| 63 |
+
if local_llm_instance:
|
| 64 |
del llm
|
| 65 |
+
del injection_vector
|
| 66 |
+
gc.collect()
|
| 67 |
+
if torch.cuda.is_available():
|
| 68 |
+
torch.cuda.empty_cache()
|
| 69 |
|
| 70 |
return results
|