Instructions to use patdev/k3-a40-bootstrap with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- llama.cpp
How to use patdev/k3-a40-bootstrap with llama.cpp:
Install (macOS, Linux)
curl -LsSf https://llama.app/install.sh | sh # Start a local OpenAI-compatible server with a web UI: llama serve -hf patdev/k3-a40-bootstrap:BF16 # Run inference directly in the terminal: llama cli -hf patdev/k3-a40-bootstrap:BF16
Install from WinGet (Windows)
winget install llama.cpp # Start a local OpenAI-compatible server with a web UI: llama serve -hf patdev/k3-a40-bootstrap:BF16 # Run inference directly in the terminal: llama cli -hf patdev/k3-a40-bootstrap:BF16
Use pre-built binary
# Download pre-built binary from: # https://github.com/ggerganov/llama.cpp/releases # Start a local OpenAI-compatible server with a web UI: ./llama-server -hf patdev/k3-a40-bootstrap:BF16 # Run inference directly in the terminal: ./llama-cli -hf patdev/k3-a40-bootstrap:BF16
Build from source code
git clone https://github.com/ggerganov/llama.cpp.git cd llama.cpp cmake -B build cmake --build build -j --target llama-server llama-cli # Start a local OpenAI-compatible server with a web UI: ./build/bin/llama-server -hf patdev/k3-a40-bootstrap:BF16 # Run inference directly in the terminal: ./build/bin/llama-cli -hf patdev/k3-a40-bootstrap:BF16
Use Docker
docker model run hf.co/patdev/k3-a40-bootstrap:BF16
- LM Studio
- Jan
- Ollama
How to use patdev/k3-a40-bootstrap with Ollama:
ollama run hf.co/patdev/k3-a40-bootstrap:BF16
- Unsloth Desktop
- Docker Model Runner
How to use patdev/k3-a40-bootstrap with Docker Model Runner:
docker model run hf.co/patdev/k3-a40-bootstrap:BF16
- Lemonade
How to use patdev/k3-a40-bootstrap with Lemonade:
Pull the model
# Download Lemonade from https://lemonade-server.ai/ lemonade pull patdev/k3-a40-bootstrap:BF16
Run and chat with the model
lemonade run user.k3-a40-bootstrap-BF16
List all available models
lemonade list
- Atomic Chat
File size: 12,901 Bytes
dbe924d | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 | """Balayage des optimisations vLLM : prefill, cache de prefixe, decodage.
Ce que la campagne a etabli, et qui dicte le choix des leviers testes ici :
le plafond solo n'est ni la bande passante (rendement 69 % sur Ada, 22,6 % sur
H200) ni le noyau MoE (`humming` = Marlin a 1 % pres sur trois architectures).
Le terme dominant est un COUT FIXE PAR JETON. Les leviers qui peuvent le
reduire sont donc : la compilation, la couverture des graphes CUDA,
l'ordonnancement, et le noyau LINEAIRE -- que `--moe-backend` ne touche pas.
Detail decisif releve dans les journaux precedents : meme avec
`--moe-backend humming`, vLLM affiche toujours
`Using MarlinNvFp4LinearKernel for NVFP4 GEMM`. Le drapeau ne remplace que les
experts. Or le chemin dense pese 1,849 Gio des 2,880 Gio du socle actif, soit
64 %. On n'avait donc echange que 36 % du travail.
Trois metriques par configuration, et non une :
- PREFILL : jetons de prompt / TTFT sur un prompt froid, abscisse lue dans
usage.prompt_tokens et jamais estimee ;
- CACHE : taux de reussite du cache de prefixe, lu dans /metrics, plus le
TTFT du meme prompt rejoue ;
- DECODAGE : debit par flux a 1 session et agrege a 8.
"""
import json
import os
import re
import statistics
import subprocess
import sys
import threading
import time
import urllib.request
MODEL = "nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4"
PORT = 8000
URL = "http://127.0.0.1:%d" % PORT
def dire(*a):
print(*a, flush=True)
def titre(t):
dire("\n" + "=" * 74)
dire(t)
dire("=" * 74)
titre("0. la carte, la pile, et la selection du noyau LINEAIRE")
subprocess.run(["nvidia-smi", "--query-gpu=name,memory.total,compute_cap",
"--format=csv,noheader"], check=False)
subprocess.run(["python3", "-c",
"import vllm,torch;print('vllm',vllm.__version__,'torch',torch.__version__,"
"'cap',torch.cuda.get_device_capability(0))"], check=False)
# Sonde gratuite : existe-t-il un moyen de changer le noyau NVFP4 LINEAIRE ?
# `--moe-backend` ne le touche pas, et c'est lui qui porte 64 % du socle.
dire("\n--- noyaux NVFP4 lineaires enregistres dans vLLM ---")
sonde = r'''
import inspect, os, re
try:
from vllm.model_executor.layers.quantization.kernels.mixed_precision import __init__ as _
except Exception:
pass
trouve = []
import vllm, pathlib
racine = pathlib.Path(vllm.__file__).parent
for p in racine.rglob("*.py"):
try:
t = p.read_text(errors="ignore")
except Exception:
continue
if "NvFp4LinearKernel" in t or "NVFP4 GEMM" in t:
for m in re.finditer(r"class\s+(\w*NvFp4\w*Kernel)\b", t):
trouve.append((m.group(1), str(p.relative_to(racine))))
for m in re.finditer(r'VLLM_[A-Z0-9_]*(?:NVFP4|GEMM|LINEAR)[A-Z0-9_]*', t):
trouve.append(("env:" + m.group(0), str(p.relative_to(racine))))
vus = set()
for nom, ou in trouve:
if nom in vus:
continue
vus.add(nom)
print(" %-42s %s" % (nom, ou))
if not vus:
print(" aucun -- la selection est probablement en dur")
'''
subprocess.run(["python3", "-c", sonde], check=False)
# Recette NVIDIA comme socle commun ; chaque essai n'ajoute que sa variante.
BASE = ["vllm", "serve", MODEL,
"--served-model-name", "ornith",
"--host", "127.0.0.1", "--port", str(PORT),
"--trust-remote-code",
"--max-model-len", "131072",
"--moe-backend", "marlin",
"--kv-cache-dtype", "fp8",
"--enable-prefix-caching",
"--gpu-memory-utilization", "0.85",
"--mamba-backend", "flashinfer",
"--mamba-cache-mode", "align",
"--reasoning-parser", "nemotron_v3",
"--tool-call-parser", "qwen3_coder",
"--enable-auto-tool-choice"]
# Un levier par ligne, pour qu'un gain soit attribuable. Le dernier essai
# combine les gagnants -- et c'est le seul dont le resultat n'est pas
# attribuable a un levier unique.
ESSAIS = [
("reference (recette NVIDIA)", [], {}),
("ordonnancement asynchrone", ["--async-scheduling"], {}),
("compilation -O3", ["-O3"], {}),
("graphes CUDA FULL", ["--compilation-config", '{"cudagraph_mode":"FULL"}'], {}),
("attention TRITON_ATTN", ["--attention-backend", "TRITON_ATTN"], {}),
("mamba-cache-mode all", ["--mamba-cache-mode", "all"], {}),
("KV en auto (pas fp8)", ["--kv-cache-dtype", "auto"], {}),
("max-num-seqs 8 (mono-session)", ["--max-num-seqs", "8"], {}),
("max-num-batched-tokens 8192", ["--max-num-batched-tokens", "8192"], {}),
("sans cache de prefixe", ["--no-enable-prefix-caching"], {}),
("echantillonneur flashinfer coupe", [], {"VLLM_USE_FLASHINFER_SAMPLER": "0"}),
]
LONG = ("Voici un module Python a auditer.\n\n" +
"\n".join("def f%d(x):\n # etape %d du pipeline de traitement\n"
" y = x * %d + %d\n return y if y > 0 else -y\n" % (i, i, i % 7 + 1, i)
for i in range(1400)))
def demarrer(sup, env_sup, journal):
env = dict(os.environ)
env.update(env_sup)
with open(journal, "w") as f:
p = subprocess.Popen(BASE + sup, stdout=f, stderr=subprocess.STDOUT, env=env)
for i in range(80):
try:
urllib.request.urlopen(URL + "/v1/models", timeout=5).read()
return p, i * 10
except Exception:
pass
if p.poll() is not None:
return None, i * 10
time.sleep(10)
p.terminate()
return None, 800
def appel(contenu, sortie=300, stream=True):
corps = json.dumps({"model": "ornith",
"messages": [{"role": "user", "content": contenu}],
"max_tokens": sortie, "temperature": 0.0,
"stream": stream,
"stream_options": {"include_usage": True} if stream else None
}).encode()
r = urllib.request.Request(URL + "/v1/chat/completions", data=corps,
headers={"Content-Type": "application/json"})
t0 = time.time()
t1 = None
n = 0
bouts = []
prompt_tokens = None
with urllib.request.urlopen(r, timeout=900) as rep:
for l in rep:
l = l.strip()
if not l.startswith(b"data: ") or l[6:] == b"[DONE]":
continue
d = json.loads(l[6:])
if d.get("usage"):
prompt_tokens = d["usage"].get("prompt_tokens")
ch = (d.get("choices") or [{}])
if not ch:
continue
de = ch[0].get("delta", {}) or {}
x = de.get("content") or de.get("reasoning") or de.get("reasoning_content")
if x:
if t1 is None:
t1 = time.time()
n += 1
bouts.append(x)
return {"ttft": (t1 - t0) if t1 else None, "n": n,
"t1": t1, "t2": time.time(), "prompt_tokens": prompt_tokens,
"txt": "".join(bouts)}
def metriques():
try:
t = urllib.request.urlopen(URL + "/metrics", timeout=15).read().decode()
except Exception:
return {}
out = {}
for cle in ("prefix_cache_queries_total", "prefix_cache_hits_total"):
m = re.search(r"vllm:gpu_%s\{[^}]*\}\s+([0-9.e+]+)" % cle, t) or \
re.search(r"vllm:%s\{[^}]*\}\s+([0-9.e+]+)" % cle, t)
if m:
out[cle] = float(m.group(1))
return out
def div4(t):
m = t.split()
if len(m) < 40:
return 1.0
g = [tuple(m[i:i + 4]) for i in range(len(m) - 3)]
return len(set(g)) / len(g)
SUJETS = ["un cache LRU avec dict et liste doublement chainee",
"un pool de connexions avec expiration et sante des sockets",
"un analyseur d'expressions par descente recursive",
"une file de priorite par tas binaire",
"un limiteur de debit par seau a jetons",
"un index inverse pour recherche plein texte",
"un ordonnanceur de taches avec dependances",
"un serialiseur binaire versionne"]
def decodage(conc):
res = [None] * conc
def un(i):
try:
res[i] = appel("Ecris en Python %s, avec trois tests unittest."
% SUJETS[i % len(SUJETS)])
except Exception as e:
res[i] = {"err": str(e)[:60]}
d0 = time.time()
fils = [threading.Thread(target=un, args=(i,)) for i in range(conc)]
for f in fils:
f.start()
for f in fils:
f.join()
d1 = time.time()
bons = [r for r in res if r and not r.get("err") and r.get("t1")]
if not bons:
return None
return {"agrege": sum(r["n"] for r in bons) / (d1 - d0),
"par_flux": statistics.median([(r["n"] - 1) / (r["t2"] - r["t1"])
for r in bons if r["t2"] > r["t1"]]),
"div4": statistics.median([div4(r["txt"]) for r in bons])}
resume = []
for idx, (etiq, sup, env_sup) in enumerate(ESSAIS):
titre("%d. %s" % (idx + 1, etiq))
journal = "/tmp/opt_%d.log" % idx
proc, secondes = demarrer(sup, env_sup, journal)
texte = open(journal, errors="replace").read()
for motif in ("NvFp4 MoE backend", "NVFP4 GEMM", "GPU KV cache size",
"attention backend", "Capturing", "cudagraph"):
for ligne in texte.splitlines():
if motif in ligne:
dire(" " + ligne.split("] ")[-1][:145])
break
if not proc:
dire(" NE DEMARRE PAS (%d s)" % secondes)
vu = set()
for ligne in texte.splitlines():
if any(m in ligne for m in ("RuntimeError", "ValueError", "Traceback",
"unrecognized arguments", "invalid choice",
"NotImplementedError", "AssertionError")):
t = ligne.split("] ")[-1][:160]
if t not in vu:
vu.add(t)
dire(" > " + t)
if len(vu) >= 4:
break
resume.append((etiq, None, None, None, None))
continue
dire(" PRET en %d s" % secondes)
# --- PREFILL, sur un prompt froid, abscisse LUE ---
m0 = metriques()
try:
froid = appel(LONG, sortie=16)
pt = froid["prompt_tokens"]
pf = (pt / froid["ttft"]) if (pt and froid["ttft"]) else None
dire(" prefill froid : %s jetons en %.2f s -> %s jetons/s"
% (pt, froid["ttft"] or 0, ("%.0f" % pf) if pf else "?"))
except Exception as e:
pt = pf = None
froid = {"ttft": None}
dire(" prefill froid : ECHEC %s" % str(e)[:70])
# --- CACHE DE PREFIXE : le meme prompt rejoue ---
gain = None
try:
chaud = appel(LONG, sortie=16)
if froid.get("ttft") and chaud.get("ttft"):
gain = froid["ttft"] / chaud["ttft"]
dire(" prefill chaud : %.2f s (x%.1f plus rapide)"
% (chaud["ttft"], gain))
except Exception as e:
dire(" prefill chaud : ECHEC %s" % str(e)[:70])
m1 = metriques()
taux = None
if m1.get("prefix_cache_queries_total") and m0 is not None:
dq = m1.get("prefix_cache_queries_total", 0) - m0.get("prefix_cache_queries_total", 0)
dh = m1.get("prefix_cache_hits_total", 0) - m0.get("prefix_cache_hits_total", 0)
if dq > 0:
taux = dh / dq
dire(" cache de prefixe : %.1f %% de reussite (%.0f/%.0f jetons)"
% (taux * 100, dh, dq))
if taux is None:
dire(" cache de prefixe : compteurs absents")
# --- DECODAGE ---
dire("conc | agrege | par flux | 4-gr")
solo = agg8 = None
for conc in (1, 8):
d = decodage(conc)
if not d:
dire("%4d | ECHEC" % conc)
continue
dire("%4d | %8.1f | %8.1f | %.3f %s"
% (conc, d["agrege"], d["par_flux"], d["div4"],
"" if d["div4"] > 0.6 else " DEGENERE"))
if conc == 1:
solo = d["par_flux"]
else:
agg8 = d["agrege"]
resume.append((etiq, solo, agg8, pf, taux))
proc.terminate()
time.sleep(20)
titre("RESUME")
dire("%-34s %8s %9s %10s %8s" % ("configuration", "solo", "agrege@8", "prefill/s", "cache"))
ref = resume[0][1] if resume and resume[0][1] else None
for etiq, solo, agg8, pf, taux in resume:
d = ""
if ref and solo:
d = " %+5.1f %%" % (100 * (solo - ref) / ref)
dire("%-34s %8s %9s %10s %8s%s"
% (etiq,
("%.1f" % solo) if solo else "echec",
("%.0f" % agg8) if agg8 else "-",
("%.0f" % pf) if pf else "-",
("%.0f %%" % (taux * 100)) if taux is not None else "-",
d))
dire("\nUn levier n'est retenu que si son gain depasse la dispersion entre deux")
dire("executions identiques -- environ 2 % sur ce banc. En dessous, c'est du bruit.")
|