drjones/lyra
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

GPU reroute: GamingPC 3070 (.222) — text qwen3.5:4b, vision minicpm-v4.5:8b; CoT-stripping line parser

Browse Source
This commit is contained in:
drjones
2026-10-01 00:08:28 +00:00
parent a7e0185ca9
commit 6f67a5a361
1 changed files with 17 additions and 9 deletions
Download Patch File Download Diff File Expand all files Collapse all files

26
app.py
View File

@@ -4,9 +4,9 @@ Single-file Flask app. Deployed on Proxmox CT 172."""
import base64, json, os, random, re, requests import base64, json, os, random, re, requests
from flask import Flask, jsonify, request, render_template_string from flask import Flask, jsonify, request, render_template_string
OLLAMA_URL = os.environ.get("LYRA_OLLAMA_URL", "http://10.30.20.69:11434") OLLAMA_URL = os.environ.get("LYRA_OLLAMA_URL", "http://10.30.20.222:11434")
TEXT_MODEL = os.environ.get("LYRA_TEXT_MODEL", "qwen3.5:9b") TEXT_MODEL = os.environ.get("LYRA_TEXT_MODEL", "ornith-1.5:9b-64k")
VISION_MODEL = os.environ.get("LYRA_VISION_MODEL", "minicpm-v4.6:1b") VISION_MODEL = os.environ.get("LYRA_VISION_MODEL", "minicpm-v4.5:8b")
app = Flask(__name__) app = Flask(__name__)
@@ -106,8 +106,10 @@ SYSTEM_VISION = (
def ollama_generate(prompt, model, images=None, timeout=60): def ollama_generate(prompt, model, images=None, timeout=60):
payload = {"model": model, "prompt": prompt, "stream": False, "think": False, payload = {"model": model, "prompt": prompt, "stream": False,
"options": {"temperature": 0.9, "num_predict": 400}} "options": {"temperature": 0.9, "num_predict": 700}}
if "qwen3" in model:
payload["think"] = False
if images: if images:
payload = {"model": model, "messages": [ payload = {"model": model, "messages": [
{"role": "user", "content": prompt, "images": images}], {"role": "user", "content": prompt, "images": images}],
@@ -116,6 +118,9 @@ def ollama_generate(prompt, model, images=None, timeout=60):
r = requests.post(f"{OLLAMA_URL}/api/chat", json=payload, timeout=timeout) r = requests.post(f"{OLLAMA_URL}/api/chat", json=payload, timeout=timeout)
r.raise_for_status() r.raise_for_status()
return (r.json().get("message") or {}).get("content", "").strip() return (r.json().get("message") or {}).get("content", "").strip()
r = requests.post(f"{OLLAMA_URL}/api/generate", json=payload, timeout=timeout)
r.raise_for_status()
return r.json().get("response", "").strip()
@app.get("/health") @app.get("/health")
@@ -130,9 +135,12 @@ def api_lines():
count = min(int(request.args.get("count", 5)), 10) count = min(int(request.args.get("count", 5)), 10)
fallback = random.sample(LINES.get(cat, LINES["witty"]), min(count, len(LINES.get(cat, [])))) fallback = random.sample(LINES.get(cat, LINES["witty"]), min(count, len(LINES.get(cat, []))))
try: try:
out = ollama_generate(SYSTEM_LINE % (count, cat, spice), TEXT_MODEL, timeout=45) out = ollama_generate(SYSTEM_LINE % (count, cat, spice), TEXT_MODEL, timeout=90)
out = re.sub(r"<think>.*?</think>", "", out, flags=re.S) # strip CoT blocks
lines = [l.strip().strip('"-•') for l in out.splitlines() if l.strip()] lines = [l.strip().strip('"-•') for l in out.splitlines() if l.strip()]
lines = [l for l in lines if 3 < len(l) < 140][:count] lines = [l for l in lines if 3 < len(l) < 140 and not re.match(
r"^(okay|let me|i (will|'ll|'m)|sure|here|the user|we need|first|next|okay so)", l, re.I)]
lines = lines[-count:] if len(lines) > count else lines # CoT models put answers last
if not lines: if not lines:
raise ValueError("empty") raise ValueError("empty")
return jsonify(source="ollama", lines=lines) return jsonify(source="ollama", lines=lines)
@@ -502,8 +510,8 @@ function timerReset(){clearInterval(tInt);clearInterval(bInt);tInt=null;$('timer
function loadSettings(){ function loadSettings(){
$('setEngine').value=S.engine||'browser';$('setVoice').value=S.voice||$('setVoice').value; $('setEngine').value=S.engine||'browser';$('setVoice').value=S.voice||$('setVoice').value;
$('setRate').value=S.rate||1;$('setPitch').value=S.pitch||1; $('setRate').value=S.rate||1;$('setPitch').value=S.pitch||1;
$('setTModel').value=S.tmodel||'qwen3.5:9b';$('setVModel').value=S.vmodel||'minicpm-v4.6:latest'; $('setTModel').value=S.tmodel||'qwen3.5:4b';$('setVModel').value=S.vmodel||'minicpm-v4.5:8b';
$('setHost').value=S.host||'http://10.30.20.69:11434';$('setAutoEvery').value=S.autoEvery||45; $('setHost').value=S.host||'http://10.30.20.222:11434';$('setAutoEvery').value=S.autoEvery||45;
$('setWarm').value=S.warm||50;$('autoMode').checked=!!S.autoMode;$('autoSpeak').checked=!!S.autoSpeak; $('setWarm').value=S.warm||50;$('autoMode').checked=!!S.autoMode;$('autoSpeak').checked=!!S.autoSpeak;
$('rateV').textContent=$('setRate').value;$('pitchV').textContent=$('setPitch').value;$('aeV').textContent=$('setAutoEvery').value; $('rateV').textContent=$('setRate').value;$('pitchV').textContent=$('setPitch').value;$('aeV').textContent=$('setAutoEvery').value;
applyWarm();} applyWarm();}