Kronos

Exact model and requirements

Artifact:

NeoQuasar (official upstream Kronos model-zoo publisher) / Kronos-mini predictor; mandatory matching Kronos-Tokenizer-2k auxiliary weights · checkpoint NeoQuasar/Kronos-mini

Publisher revision:

f4e68697d9d5aed55cef5c96aabc3376bcad9f81

Artifact digest:

sha256:a7d5f37e2e9fbd9891f7d7d4f72574512dd1f704fee14223e0a8cd0fbf54197c

Download size:

0.02 GiB

Memory requirement (unknown):

Unknown — No measured peak for both networks, CPU tensors, native decoding, dependencies and serialization.

System RAM / disk:

Unknown / Unknown

Placement:

single device. One CPU host; no model sharding. Both networks and all input/output processing remain on CPU; no GPU or disk offload.

KV cache:

Pinned native inference recomputes the paired-code sequence window each step; no autoregressive K/V cache is configured.

Workload:

128 token context; 64 input tokens; 8 output tokens; concurrency 1; batch 1; cache unknown; repetitions Unknown.

Tokenizer:

Native learned Kronos-Tokenizer-2k; two 10-bit scalar code streams per bar position · 26966d0035065a0cae0ebad7af8ece35bc1fb51c

Sampling / reasoning:

Fixed seed 20251016 reset per request; temperature 1.0, top_k 0, top_p 0.9, sample_count 1. Native inference samples coarse then fine IDs for every future bar. Eight forecast positions mean 16 scalar IDs per path, not eight scalar IDs or text tokens. Seed does not guarantee cross-version or cross-hardware identity. No chat, prompt template, language reasoning or tool calling. Context 128, input 64 and output eight denote paired-code autoregressive bar positions. Both streams are embedded into one position. Mini published context is 2048; this author workload is a reduced window, not a new global model maximum. Total 72 positions never reach the rolling-window boundary. half=True in native encoding splits hierarchical codes; execution remains float32.

Runtime prerequisites:

Ubuntu 24.04 x86_64; See prerequisites above.

Official artifact · License

Upstream source MIT license and both official artifact cards declare MIT. Retain the copyright and permission notice in copies or substantial portions; supplied without warranty. This original synthetic fixture has no external market-data dependency. MIT does not grant rights to third-party financial datasets or provide trading, accuracy or legal assurances.

• Mini architecture config · 0 GiB · revision f4e68697d9d5aed55cef5c96aabc3376bcad9f81
• Mandatory learned Tokenizer-2k weights · 0.01 GiB · revision 26966d0035065a0cae0ebad7af8ece35bc1fb51c
• Tokenizer-2k architecture config · 0 GiB · revision 26966d0035065a0cae0ebad7af8ece35bc1fb51c
• MIT notice · 0 GiB · revision 67b630e67f6a18c9e9be918d9b4337c960db1e9a

• Main predictor LFS artifact is 16,440,776 bytes; mandatory learned tokenizer adds 15,842,376 bytes. Their 32,283,152-byte total is serialized weight metadata, not peak process memory. No weight body downloaded or independently hashed by author.
• Four CPU threads, 32 GiB RAM and 40 GiB disk are author allocations. Required CPU memory, fit and performance remain unknown.

Install the runtime

1. 1. Create the pinned native CPU research environment

   Use a dedicated directory, the official hashed CPU wheel and the exact upstream commit. The selected direct dependency versions are source-qualified author choices; no CUDA driver or Transformers runtime is needed. Record resolved transitive and system versions.

   CommandCopy command

   sudo apt-get update
   sudo apt-get install -y python3.12 python3.12-venv git curl ca-certificates openssl
   mkdir -p "$HOME/kronos-mini-forecast/checkpoints/model" "$HOME/kronos-mini-forecast/checkpoints/tokenizer"
   cd "$HOME/kronos-mini-forecast"
   python3.12 -m venv .venv
   .venv/bin/python -m pip install 'https://download.pytorch.org/whl/cpu/torch-2.6.0%2Bcpu-cp312-cp312-linux_x86_64.whl#sha256=59e78aa0c690f70734e42670036d6b541930b8eabbaa18d94e090abf14cc4d91'
   .venv/bin/python -m pip install 'numpy==1.26.4' 'pandas==2.2.2' 'einops==0.8.1' 'huggingface_hub==0.33.1' 'safetensors==0.6.2' 'tqdm==4.67.1'
   .venv/bin/python -m pip check
   .venv/bin/python -m pip freeze > resolved-environment.txt
   git clone https://github.com/shiyu-coder/Kronos.git runtime
   git -C runtime checkout --detach 67b630e67f6a18c9e9be918d9b4337c960db1e9a
   test "$(git -C runtime rev-parse HEAD)" = '67b630e67f6a18c9e9be918d9b4337c960db1e9a'

   Expected result: pip check succeeds, environment is recorded, and the runtime checkout matches the pinned commit. This has not been installed by the author.

Download the model

1. 1. Retrieve both exact native checkpoints and notice

   Read the MIT notice first. Download the mini predictor and its mandatory matching Tokenizer-2k, not the base tokenizer. Verify all five file hashes. Weight hashes and sizes are official LFS metadata; author did not retrieve weights.

   CommandCopy command

   cd "$HOME/kronos-mini-forecast"
   curl --fail --location --output checkpoints/model/model.safetensors 'https://huggingface.co/NeoQuasar/Kronos-mini/resolve/f4e68697d9d5aed55cef5c96aabc3376bcad9f81/model.safetensors'
   curl --fail --location --output checkpoints/model/config.json 'https://huggingface.co/NeoQuasar/Kronos-mini/resolve/f4e68697d9d5aed55cef5c96aabc3376bcad9f81/config.json'
   curl --fail --location --output checkpoints/tokenizer/model.safetensors 'https://huggingface.co/NeoQuasar/Kronos-Tokenizer-2k/resolve/26966d0035065a0cae0ebad7af8ece35bc1fb51c/model.safetensors'
   curl --fail --location --output checkpoints/tokenizer/config.json 'https://huggingface.co/NeoQuasar/Kronos-Tokenizer-2k/resolve/26966d0035065a0cae0ebad7af8ece35bc1fb51c/config.json'
   curl --fail --location --output KRONOS-LICENSE 'https://raw.githubusercontent.com/shiyu-coder/Kronos/67b630e67f6a18c9e9be918d9b4337c960db1e9a/LICENSE'
   sha256sum --check <<'SHA256'
   a7d5f37e2e9fbd9891f7d7d4f72574512dd1f704fee14223e0a8cd0fbf54197c  checkpoints/model/model.safetensors
   70daca2cb11e3a979dd6b8ac12ee08e2aace877acf28f5b8dfb4fe5609736201  checkpoints/model/config.json
   b97ec46b3b72160509e289183eaf7bdf5f0dac5bb9b49522f6d46638a99a8717  checkpoints/tokenizer/model.safetensors
   0b30a443affb03e05a876a083857de9164f899feb7b4d261da02c485c9a3e3b6  checkpoints/tokenizer/config.json
   acb2d194d378204e5f2be4dcd24d39ecac437903620c790c3315a96dab388fdc  KRONOS-LICENSE
   SHA256

   Expected result: Both checkpoint directories contain verified configs and native safetensors; exact MIT notice retained.

Configure

1. 1. Write the bounded private research wrapper

   This reader-authored wrapper is not publisher serving code. Native checkpoint loading is strict; both networks are explicitly CPU float32 in evaluation mode. Only a fixed synthetic fixture is accepted. Body framing and JSON errors are rejected before inference. This serial research server has a socket read timeout, no TLS, no inference deadline and no public-service hardening.

   CommandCopy command

   cd "$HOME/kronos-mini-forecast"
   cat > reader-forecast-server.py <<'PYWRAPPER'
   """Reader-authored fixed synthetic Kronos research endpoint; not publisher serving code."""
   import datetime as dt
   import hmac
   import json
   import math
   import os
   from http.server import BaseHTTPRequestHandler, HTTPServer
   from pathlib import Path
   
   ROOT = Path(__file__).resolve().parent
   COLUMNS = ['open', 'high', 'low', 'close', 'volume', 'amount']
   SEED = 20251016
   
   def fixture():
       start = dt.datetime(2025, 1, 2, 9, 30)
       times = [(start + dt.timedelta(minutes=5*i)).isoformat() for i in range(72)]
       rows = [[100+i, 102+i, 99+i, 101+i, 1000+10*i, (1000+10*i)*(201+2*i)/2] for i in range(64)]
       return dict(columns=COLUMNS, history=rows, history_timestamps=times[:64], future_timestamps=times[64:])
   
   def result(values):
       if not isinstance(values, list) or len(values) != 8:
           raise ValueError('forecast shape')
       for row in values:
           if not isinstance(row, list) or len(row) != 6:
               raise ValueError('forecast shape')
           if any(isinstance(v, bool) or not isinstance(v, (float, int)) or not math.isfinite(v) for v in row):
               raise ValueError('nonfinite forecast')
       return dict(fixture='synthetic-ohlcva-v1', columns=COLUMNS,
                   timestamps=fixture()['future_timestamps'], forecast=values,
                   history_positions=64, forecast_positions=8, context_positions=128,
                   scalar_codes_per_position=2, generated_scalar_ids_per_path=16,
                   sample_count=1, seed=SEED, generated_text_tokens=None)
   
   def object_pairs(pairs):
       out = {}
       for k, v in pairs:
           if k in out:
               raise ValueError('duplicate key')
           out[k] = v
       return out
   
   def create_server(predict, key, port=8993):
       if len(key) < 32 or not key.isascii() or not key.isalnum():
           raise ValueError('key must be at least 32 ASCII alphanumeric characters')
       class Handler(BaseHTTPRequestHandler):
           protocol_version = 'HTTP/1.0'
           def setup(self):
               super().setup()
               self.connection.settimeout(5)
           def log_message(self, *_):
               pass
           def reply(self, code, obj):
               data = json.dumps(obj, allow_nan=False, separators=(',', ':')).encode()
               self.send_response(code)
               self.send_header('Content-Type', 'application/json')
               self.send_header('Content-Length', str(len(data)))
               self.send_header('Connection', 'close')
               self.end_headers()
               try:
                   self.wfile.write(data)
               except OSError:
                   pass
           def authorized(self):
               vals = self.headers.get_all('Authorization') or []
               if len(vals) != 1 or not hmac.compare_digest(vals[0].encode(), ('Bearer '+key).encode()):
                   self.reply(401, {'error': 'unauthorized'})
                   return False
               return True
           def do_GET(self):
               if not self.authorized(): return
               if self.path != '/health':
                   self.reply(404, {'error': 'not found'}); return
               self.reply(200, dict(status='ready', model='Kronos-mini', tokenizer='Kronos-Tokenizer-2k', device='cpu', dtype='float32', context_positions=128, history_positions=64, forecast_positions=8))
           def do_POST(self):
               if not self.authorized(): return
               if self.path != '/v1/forecast':
                   self.reply(404, {'error': 'not found'}); return
               if self.headers.get_all('Transfer-Encoding'):
                   self.reply(400, {'error': 'transfer encoding unsupported'}); return
               lens = self.headers.get_all('Content-Length') or []
               if len(lens) != 1 or not lens[0] or not lens[0].isascii() or not lens[0].isdigit():
                   self.reply(411, {'error': 'one ASCII content length required'}); return
               if len(lens[0]) > 4 or int(lens[0]) > 256:
                   self.reply(413, {'error': 'body too large'}); return
               if (self.headers.get_all('Content-Type') or []) != ['application/json']:
                   self.reply(415, {'error': 'application/json required'}); return
               try:
                   raw = self.rfile.read(int(lens[0]))
                   if len(raw) != int(lens[0]): raise ValueError('short body')
                   obj = json.loads(raw, object_pairs_hook=object_pairs, parse_constant=lambda _: (_ for _ in ()).throw(ValueError('nonfinite')))
                   if obj != {'fixture': 'synthetic-ohlcva-v1'}: raise ValueError('fixed fixture only')
               except (ValueError, UnicodeError, OSError):
                   self.reply(400, {'error': 'invalid fixed-fixture request'}); return
               try:
                   payload = result(predict(fixture()))
               except Exception:
                   self.reply(500, {'error': 'forecast failed'}); return
               self.reply(200, payload)
       return HTTPServer(('127.0.0.1', port), Handler)
   
   def main():
       import sys
       sys.path.insert(0, str(ROOT/'runtime'))
       import torch
       import pandas as pd
       from model import Kronos, KronosTokenizer, KronosPredictor
       torch.set_num_threads(4)
       cfg = json.loads((ROOT/'checkpoints/model/config.json').read_text())
       tcfg = json.loads((ROOT/'checkpoints/tokenizer/config.json').read_text())
       assert (cfg['d_model'], cfg['n_layers'], cfg['n_heads'], cfg['s1_bits'], cfg['s2_bits']) == (256, 4, 4, 10, 10)
       assert (tcfg['d_in'], tcfg['group_size'], tcfg['s1_bits'], tcfg['s2_bits']) == (6, 5, 10, 10)
       model = Kronos.from_pretrained(str(ROOT/'checkpoints/model'), map_location='cpu', strict=True, local_files_only=True).float().cpu().eval()
       tokenizer = KronosTokenizer.from_pretrained(str(ROOT/'checkpoints/tokenizer'), map_location='cpu', strict=True, local_files_only=True).float().cpu().eval()
       assert not model.training and not tokenizer.training
       assert all(p.device.type == 'cpu' and p.dtype == torch.float32 for m in (model, tokenizer) for p in m.parameters())
       predictor = KronosPredictor(model, tokenizer, device='cpu', max_context=128, clip=5)
       expected = fixture()
       (ROOT/'synthetic-fixture.json').write_text(json.dumps(expected, indent=2)+'\n')
       def predict(f):
           assert f == expected and len(f['history']) == 64 and len(f['future_timestamps']) == 8
           assert 64+8 <= 128 and all(math.isfinite(v) for r in f['history'] for v in r)
           torch.manual_seed(SEED)
           df = pd.DataFrame(f['history'], columns=COLUMNS)
           x = pd.Series(pd.to_datetime(f['history_timestamps']))
           y = pd.Series(pd.to_datetime(f['future_timestamps']))
           out = predictor.predict(df, x, y, pred_len=8, T=1.0, top_k=0, top_p=0.9, sample_count=1, verbose=False)
           assert list(out.columns) == COLUMNS and out.shape == (8, 6)
           assert list(out.index) == list(y)
           return out.to_numpy().tolist()
       key = (ROOT/'server.key').read_text().strip()
       create_server(predict, key).serve_forever()
   
   if __name__ == '__main__':
       main()
   PYWRAPPER
   sha256sum --check <<'SHA256'
   2251375140e107d89d5e60485f548b8574f144cf4c323d3134c532c0677e3865  reader-forecast-server.py
   SHA256

   Expected result: Wrapper hash matches. Startup will fail rather than silently accept incompatible state dictionaries or training-mode execution.

Start the model

1. 1. Load both networks and serve on loopback

   Run in the foreground; keep the terminal open. Preserve the private key across restarts and never paste it into logs. Health becomes available only after successful strict model/tokenizer loading. Use another terminal for checks.

   CommandCopy command

   cd "$HOME/kronos-mini-forecast"
   umask 077
   if [ ! -e server.key ]; then openssl rand -hex 32 > server.key; fi
   test -s server.key
   chmod 600 server.key
   OMP_NUM_THREADS=4 MKL_NUM_THREADS=4 HF_HUB_OFFLINE=1 .venv/bin/python reader-forecast-server.py

   Expected result: Native predictor and tokenizer load locally, fixture is written, and the private serial service binds only to loopback port 8993. Startup time and fit are unknown.

Verify the setup

1. 1. Check authenticated readiness

   In a second server terminal use the dedicated virtual environment explicitly. Do not print the key.

   CommandCopy command

   cd "$HOME/kronos-mini-forecast"
   KRONOS_KEY=$(cat server.key)
   curl --fail --silent --show-error -H "Authorization: Bearer $KRONOS_KEY" http://127.0.0.1:8993/health > health.json
   .venv/bin/python - <<'PYHEALTH'
   import json
   x=json.load(open('health.json'))
   assert x['status']=='ready' and x['device']=='cpu' and x['dtype']=='float32'
   assert (x['context_positions'],x['history_positions'],x['forecast_positions'])==(128,64,8)
   print('ready fixed synthetic forecast endpoint')
   PYHEALTH

   Expected result: Ready response reports CPU float32, 128 configured paired-code positions, 64 history and eight forecast positions.

2. 2. Forecast the original fixed synthetic sequence

   Submit exactly one fixture selector. The fixture contains 64 original five-minute OHLCVA rows from a synthetic timezone-naive calendar; it is not an exchange calendar. The request does not accept parameters or arbitrary data. Verify fixture bytes, shape, future timestamps, finite numbers and explicit native units. No exact forecast values or quality threshold are asserted.

   CommandCopy command

   cd "$HOME/kronos-mini-forecast"
   KRONOS_KEY=$(cat server.key)
   curl --fail --silent --show-error --max-time 600 -H "Authorization: Bearer $KRONOS_KEY" -H 'Content-Type: application/json' --data '{"fixture":"synthetic-ohlcva-v1"}' http://127.0.0.1:8993/v1/forecast > forecast.json
   .venv/bin/python - <<'PYCHECK'
   import json,math,datetime,hashlib
   assert hashlib.sha256(open('synthetic-fixture.json','rb').read()).hexdigest()=='f532d0bd540cabb9fabd8107af42986a83f62d5246a025f6bd657c6217f74506'
   x=json.load(open('forecast.json'))
   assert x['columns']==['open','high','low','close','volume','amount']
   assert (x['history_positions'],x['forecast_positions'],x['context_positions'])==(64,8,128)
   assert x['scalar_codes_per_position']==2 and x['generated_scalar_ids_per_path']==16
   assert x['generated_text_tokens'] is None and x['sample_count']==1 and x['seed']==20251016
   start=datetime.datetime(2025,1,2,9,30)
   assert x['timestamps']==[(start+datetime.timedelta(minutes=5*i)).isoformat() for i in range(64,72)]
   assert len(x['forecast'])==8
   assert all(len(r)==6 and all(type(v) in (int,float) and math.isfinite(v) for v in r) for r in x['forecast'])
   print('eight finite six-value forecast rows, exact future timestamps and explicit paired-code units')
   PYCHECK

   Expected result: Eight finite rows with six named values and exact future timestamps pass the checks. A client timeout is not cancellation; stop the foreground process if needed. This expected behavior was not model-executed by the author.

3. 3. Verify the private access boundary

   Health requires the same Bearer key as forecasting. An unauthenticated request must fail.

   CommandCopy command

   cd "$HOME/kronos-mini-forecast"
   test "$(curl --silent --output unauthorized.json --write-out '%{http_code}' http://127.0.0.1:8993/health)" = 401
   .venv/bin/python - <<'PYAUTH'
   import json
   assert json.load(open('unauthorized.json'))=={'error':'unauthorized'}
   print('unauthenticated health request rejected')
   PYAUTH

   Expected result: HTTP 401 and structured unauthorized JSON; no model work is triggered.

Performance evidence

No compatible measurement available. Documented support and estimated memory do not establish response speed.

Limits and troubleshooting

• The custom wrapper is an educational private serial endpoint, not an official or production server. It accepts only the original fixed synthetic fixture; no arbitrary dataset, file path, network URL, market symbol or variable horizon is supported.
• The unit in context, input and output fields is one autoregressive bar position containing two scalar 10-bit IDs. Each stream has 1024 possible IDs. Eight output positions correspond to 16 scalar generated IDs per sampled path and decode to eight rows of six values, not text.
• The configured 128-position window is deliberately below mini published 2048-position context. Sixty-four history plus eight future positions fit without rollover. Native predictor encodes history before window cropping; this wrapper bounds input before calling it and does not silently truncate. Small/base checkpoints and the unreleased large variant remain separate unresolved jobs.
• Both networks explicitly use eval mode and CPU float32. Source Hub safetensors loading does not itself call eval, and native predictor only moves devices. Strict checkpoint loading and saved-config assertions are required. No model execution, peak-memory or latency measurement was performed.
• Native preprocessing computes mean and standard deviation over provided history, normalizes using std plus a small stabilizer, clips normalized values to minus five through five, and inverse-scales predictions. All six columns are supplied to avoid missing-volume/amount defaults. Minute, hour, weekday, day and month are encoded; year, seconds, timezone and exchange calendars are not.
• Native decoded output is unconstrained: prices or volumes may be negative, and OHLC relationships need not hold. Finite shape checks are interface checks, not forecast accuracy, market validity, safety, investment advice or profit guarantees. No real financial data or trading is used.
• Fixed seed aids investigation but is not a guarantee of identical numerical results across runtime versions or hardware. Client timeout does not cancel in-progress CPU work; the server has no generation deadline. Keep it on trusted loopback.

Checkpoint or config mismatch

If strict loading or saved-config assertions fail, stop and verify the exact mini predictor and matching Tokenizer-2k snapshots and source checkout. Do not weaken strict loading or substitute the base tokenizer.

Invalid forecast result

If results contain nonfinite values or the expected eight-by-six shape/index fails, the wrapper returns a generic failure; retain the pinned environment and investigate offline. Do not treat shape success as a quality result.

Insufficient resources

If startup exhausts memory or is too slow, stop the process. Allocated RAM and threads are not measured fit or speed evidence; no reduced-precision or remote fallback is configured.

HTTP 401

If HTTP 401 occurs, read the existing private server.key only in the client shell and send a single Bearer header. Never print, commit or upload the key.

Request timeout

If a request hangs, the 600-second client limit only stops waiting. Use Ctrl+C in the foreground server terminal to stop native work before retrying.

Stop the model

1. 1. Stop the private forecasting process

   Press Ctrl+C in the foreground server terminal. In each client shell unset the key; confirm loopback no longer serves.

   CommandCopy command

   unset KRONOS_KEY
   if curl --silent --max-time 2 http://127.0.0.1:8993/health >/dev/null; then echo "Port still responds; inspect the foreground process"; exit 1; fi

   Expected result: Foreground process stopped and loopback port no longer responds.

Uninstall

1. 1. Remove dedicated research files

   Preserve desired outputs and license notices first. Confirm the dedicated path before removal. Shared Ubuntu packages remain.

   CommandCopy command

   rm -r -- "$HOME/kronos-mini-forecast"

   Expected result: Dedicated weights, source checkout, virtual environment, wrapper, fixture, results and private key removed.

Exact model and requirements

Artifact:

NeoQuasar (official upstream Kronos model-zoo publisher) / Kronos-mini predictor; mandatory matching Kronos-Tokenizer-2k auxiliary weights · checkpoint NeoQuasar/Kronos-mini

Publisher revision:

f4e68697d9d5aed55cef5c96aabc3376bcad9f81

Artifact digest:

sha256:a7d5f37e2e9fbd9891f7d7d4f72574512dd1f704fee14223e0a8cd0fbf54197c

Download size:

0.02 GiB

Memory requirement (unknown):

Unknown — No measured peak for both networks, CPU tensors, native decoding, dependencies and serialization.

System RAM / disk:

Unknown / Unknown

Placement:

single device. One CPU host; no model sharding. Both networks and all input/output processing remain on CPU; no GPU or disk offload.

KV cache:

Pinned native inference recomputes the paired-code sequence window each step; no autoregressive K/V cache is configured.

Workload:

128 token context; 64 input tokens; 8 output tokens; concurrency 1; batch 1; cache unknown; repetitions Unknown.

Tokenizer:

Native learned Kronos-Tokenizer-2k; two 10-bit scalar code streams per bar position · 26966d0035065a0cae0ebad7af8ece35bc1fb51c

Sampling / reasoning:

Fixed seed 20251016 reset per request; temperature 1.0, top_k 0, top_p 0.9, sample_count 1. Native inference samples coarse then fine IDs for every future bar. Eight forecast positions mean 16 scalar IDs per path, not eight scalar IDs or text tokens. Seed does not guarantee cross-version or cross-hardware identity. No chat, prompt template, language reasoning or tool calling. Context 128, input 64 and output eight denote paired-code autoregressive bar positions. Both streams are embedded into one position. Mini published context is 2048; this author workload is a reduced window, not a new global model maximum. Total 72 positions never reach the rolling-window boundary. half=True in native encoding splits hierarchical codes; execution remains float32.

Runtime prerequisites:

Ubuntu 24.04 x86_64; See prerequisites above.

Official artifact · License

Upstream source MIT license and both official artifact cards declare MIT. Retain the copyright and permission notice in copies or substantial portions; supplied without warranty. This original synthetic fixture has no external market-data dependency. MIT does not grant rights to third-party financial datasets or provide trading, accuracy or legal assurances.

• Mini architecture config · 0 GiB · revision f4e68697d9d5aed55cef5c96aabc3376bcad9f81
• Mandatory learned Tokenizer-2k weights · 0.01 GiB · revision 26966d0035065a0cae0ebad7af8ece35bc1fb51c
• Tokenizer-2k architecture config · 0 GiB · revision 26966d0035065a0cae0ebad7af8ece35bc1fb51c
• MIT notice · 0 GiB · revision 67b630e67f6a18c9e9be918d9b4337c960db1e9a

• Main predictor LFS artifact is 16,440,776 bytes; mandatory learned tokenizer adds 15,842,376 bytes. Their 32,283,152-byte total is serialized weight metadata, not peak process memory. No weight body downloaded or independently hashed by author.
• Four CPU threads, 32 GiB RAM and 40 GiB disk are author allocations. Required CPU memory, fit and performance remain unknown.

Install the runtime

1. 1. Create the pinned native CPU research environment

   Use a dedicated directory, the official hashed CPU wheel and the exact upstream commit. The selected direct dependency versions are source-qualified author choices; no CUDA driver or Transformers runtime is needed. Record resolved transitive and system versions.

   CommandCopy command

   sudo apt-get update
   sudo apt-get install -y python3.12 python3.12-venv git curl ca-certificates openssl
   mkdir -p "$HOME/kronos-mini-forecast/checkpoints/model" "$HOME/kronos-mini-forecast/checkpoints/tokenizer"
   cd "$HOME/kronos-mini-forecast"
   python3.12 -m venv .venv
   .venv/bin/python -m pip install 'https://download.pytorch.org/whl/cpu/torch-2.6.0%2Bcpu-cp312-cp312-linux_x86_64.whl#sha256=59e78aa0c690f70734e42670036d6b541930b8eabbaa18d94e090abf14cc4d91'
   .venv/bin/python -m pip install 'numpy==1.26.4' 'pandas==2.2.2' 'einops==0.8.1' 'huggingface_hub==0.33.1' 'safetensors==0.6.2' 'tqdm==4.67.1'
   .venv/bin/python -m pip check
   .venv/bin/python -m pip freeze > resolved-environment.txt
   git clone https://github.com/shiyu-coder/Kronos.git runtime
   git -C runtime checkout --detach 67b630e67f6a18c9e9be918d9b4337c960db1e9a
   test "$(git -C runtime rev-parse HEAD)" = '67b630e67f6a18c9e9be918d9b4337c960db1e9a'

   Expected result: pip check succeeds, environment is recorded, and the runtime checkout matches the pinned commit. This has not been installed by the author.

Download the model

1. 1. Retrieve both exact native checkpoints and notice

   Read the MIT notice first. Download the mini predictor and its mandatory matching Tokenizer-2k, not the base tokenizer. Verify all five file hashes. Weight hashes and sizes are official LFS metadata; author did not retrieve weights.

   CommandCopy command

   cd "$HOME/kronos-mini-forecast"
   curl --fail --location --output checkpoints/model/model.safetensors 'https://huggingface.co/NeoQuasar/Kronos-mini/resolve/f4e68697d9d5aed55cef5c96aabc3376bcad9f81/model.safetensors'
   curl --fail --location --output checkpoints/model/config.json 'https://huggingface.co/NeoQuasar/Kronos-mini/resolve/f4e68697d9d5aed55cef5c96aabc3376bcad9f81/config.json'
   curl --fail --location --output checkpoints/tokenizer/model.safetensors 'https://huggingface.co/NeoQuasar/Kronos-Tokenizer-2k/resolve/26966d0035065a0cae0ebad7af8ece35bc1fb51c/model.safetensors'
   curl --fail --location --output checkpoints/tokenizer/config.json 'https://huggingface.co/NeoQuasar/Kronos-Tokenizer-2k/resolve/26966d0035065a0cae0ebad7af8ece35bc1fb51c/config.json'
   curl --fail --location --output KRONOS-LICENSE 'https://raw.githubusercontent.com/shiyu-coder/Kronos/67b630e67f6a18c9e9be918d9b4337c960db1e9a/LICENSE'
   sha256sum --check <<'SHA256'
   a7d5f37e2e9fbd9891f7d7d4f72574512dd1f704fee14223e0a8cd0fbf54197c  checkpoints/model/model.safetensors
   70daca2cb11e3a979dd6b8ac12ee08e2aace877acf28f5b8dfb4fe5609736201  checkpoints/model/config.json
   b97ec46b3b72160509e289183eaf7bdf5f0dac5bb9b49522f6d46638a99a8717  checkpoints/tokenizer/model.safetensors
   0b30a443affb03e05a876a083857de9164f899feb7b4d261da02c485c9a3e3b6  checkpoints/tokenizer/config.json
   acb2d194d378204e5f2be4dcd24d39ecac437903620c790c3315a96dab388fdc  KRONOS-LICENSE
   SHA256

   Expected result: Both checkpoint directories contain verified configs and native safetensors; exact MIT notice retained.

Configure

1. 1. Write the bounded private research wrapper

   This reader-authored wrapper is not publisher serving code. Native checkpoint loading is strict; both networks are explicitly CPU float32 in evaluation mode. Only a fixed synthetic fixture is accepted. Body framing and JSON errors are rejected before inference. This serial research server has a socket read timeout, no TLS, no inference deadline and no public-service hardening.

   CommandCopy command

   cd "$HOME/kronos-mini-forecast"
   cat > reader-forecast-server.py <<'PYWRAPPER'
   """Reader-authored fixed synthetic Kronos research endpoint; not publisher serving code."""
   import datetime as dt
   import hmac
   import json
   import math
   import os
   from http.server import BaseHTTPRequestHandler, HTTPServer
   from pathlib import Path
   
   ROOT = Path(__file__).resolve().parent
   COLUMNS = ['open', 'high', 'low', 'close', 'volume', 'amount']
   SEED = 20251016
   
   def fixture():
       start = dt.datetime(2025, 1, 2, 9, 30)
       times = [(start + dt.timedelta(minutes=5*i)).isoformat() for i in range(72)]
       rows = [[100+i, 102+i, 99+i, 101+i, 1000+10*i, (1000+10*i)*(201+2*i)/2] for i in range(64)]
       return dict(columns=COLUMNS, history=rows, history_timestamps=times[:64], future_timestamps=times[64:])
   
   def result(values):
       if not isinstance(values, list) or len(values) != 8:
           raise ValueError('forecast shape')
       for row in values:
           if not isinstance(row, list) or len(row) != 6:
               raise ValueError('forecast shape')
           if any(isinstance(v, bool) or not isinstance(v, (float, int)) or not math.isfinite(v) for v in row):
               raise ValueError('nonfinite forecast')
       return dict(fixture='synthetic-ohlcva-v1', columns=COLUMNS,
                   timestamps=fixture()['future_timestamps'], forecast=values,
                   history_positions=64, forecast_positions=8, context_positions=128,
                   scalar_codes_per_position=2, generated_scalar_ids_per_path=16,
                   sample_count=1, seed=SEED, generated_text_tokens=None)
   
   def object_pairs(pairs):
       out = {}
       for k, v in pairs:
           if k in out:
               raise ValueError('duplicate key')
           out[k] = v
       return out
   
   def create_server(predict, key, port=8993):
       if len(key) < 32 or not key.isascii() or not key.isalnum():
           raise ValueError('key must be at least 32 ASCII alphanumeric characters')
       class Handler(BaseHTTPRequestHandler):
           protocol_version = 'HTTP/1.0'
           def setup(self):
               super().setup()
               self.connection.settimeout(5)
           def log_message(self, *_):
               pass
           def reply(self, code, obj):
               data = json.dumps(obj, allow_nan=False, separators=(',', ':')).encode()
               self.send_response(code)
               self.send_header('Content-Type', 'application/json')
               self.send_header('Content-Length', str(len(data)))
               self.send_header('Connection', 'close')
               self.end_headers()
               try:
                   self.wfile.write(data)
               except OSError:
                   pass
           def authorized(self):
               vals = self.headers.get_all('Authorization') or []
               if len(vals) != 1 or not hmac.compare_digest(vals[0].encode(), ('Bearer '+key).encode()):
                   self.reply(401, {'error': 'unauthorized'})
                   return False
               return True
           def do_GET(self):
               if not self.authorized(): return
               if self.path != '/health':
                   self.reply(404, {'error': 'not found'}); return
               self.reply(200, dict(status='ready', model='Kronos-mini', tokenizer='Kronos-Tokenizer-2k', device='cpu', dtype='float32', context_positions=128, history_positions=64, forecast_positions=8))
           def do_POST(self):
               if not self.authorized(): return
               if self.path != '/v1/forecast':
                   self.reply(404, {'error': 'not found'}); return
               if self.headers.get_all('Transfer-Encoding'):
                   self.reply(400, {'error': 'transfer encoding unsupported'}); return
               lens = self.headers.get_all('Content-Length') or []
               if len(lens) != 1 or not lens[0] or not lens[0].isascii() or not lens[0].isdigit():
                   self.reply(411, {'error': 'one ASCII content length required'}); return
               if len(lens[0]) > 4 or int(lens[0]) > 256:
                   self.reply(413, {'error': 'body too large'}); return
               if (self.headers.get_all('Content-Type') or []) != ['application/json']:
                   self.reply(415, {'error': 'application/json required'}); return
               try:
                   raw = self.rfile.read(int(lens[0]))
                   if len(raw) != int(lens[0]): raise ValueError('short body')
                   obj = json.loads(raw, object_pairs_hook=object_pairs, parse_constant=lambda _: (_ for _ in ()).throw(ValueError('nonfinite')))
                   if obj != {'fixture': 'synthetic-ohlcva-v1'}: raise ValueError('fixed fixture only')
               except (ValueError, UnicodeError, OSError):
                   self.reply(400, {'error': 'invalid fixed-fixture request'}); return
               try:
                   payload = result(predict(fixture()))
               except Exception:
                   self.reply(500, {'error': 'forecast failed'}); return
               self.reply(200, payload)
       return HTTPServer(('127.0.0.1', port), Handler)
   
   def main():
       import sys
       sys.path.insert(0, str(ROOT/'runtime'))
       import torch
       import pandas as pd
       from model import Kronos, KronosTokenizer, KronosPredictor
       torch.set_num_threads(4)
       cfg = json.loads((ROOT/'checkpoints/model/config.json').read_text())
       tcfg = json.loads((ROOT/'checkpoints/tokenizer/config.json').read_text())
       assert (cfg['d_model'], cfg['n_layers'], cfg['n_heads'], cfg['s1_bits'], cfg['s2_bits']) == (256, 4, 4, 10, 10)
       assert (tcfg['d_in'], tcfg['group_size'], tcfg['s1_bits'], tcfg['s2_bits']) == (6, 5, 10, 10)
       model = Kronos.from_pretrained(str(ROOT/'checkpoints/model'), map_location='cpu', strict=True, local_files_only=True).float().cpu().eval()
       tokenizer = KronosTokenizer.from_pretrained(str(ROOT/'checkpoints/tokenizer'), map_location='cpu', strict=True, local_files_only=True).float().cpu().eval()
       assert not model.training and not tokenizer.training
       assert all(p.device.type == 'cpu' and p.dtype == torch.float32 for m in (model, tokenizer) for p in m.parameters())
       predictor = KronosPredictor(model, tokenizer, device='cpu', max_context=128, clip=5)
       expected = fixture()
       (ROOT/'synthetic-fixture.json').write_text(json.dumps(expected, indent=2)+'\n')
       def predict(f):
           assert f == expected and len(f['history']) == 64 and len(f['future_timestamps']) == 8
           assert 64+8 <= 128 and all(math.isfinite(v) for r in f['history'] for v in r)
           torch.manual_seed(SEED)
           df = pd.DataFrame(f['history'], columns=COLUMNS)
           x = pd.Series(pd.to_datetime(f['history_timestamps']))
           y = pd.Series(pd.to_datetime(f['future_timestamps']))
           out = predictor.predict(df, x, y, pred_len=8, T=1.0, top_k=0, top_p=0.9, sample_count=1, verbose=False)
           assert list(out.columns) == COLUMNS and out.shape == (8, 6)
           assert list(out.index) == list(y)
           return out.to_numpy().tolist()
       key = (ROOT/'server.key').read_text().strip()
       create_server(predict, key).serve_forever()
   
   if __name__ == '__main__':
       main()
   PYWRAPPER
   sha256sum --check <<'SHA256'
   2251375140e107d89d5e60485f548b8574f144cf4c323d3134c532c0677e3865  reader-forecast-server.py
   SHA256

   Expected result: Wrapper hash matches. Startup will fail rather than silently accept incompatible state dictionaries or training-mode execution.

Start the model

1. 1. Load both networks and serve on loopback

   Run in the foreground; keep the terminal open. Preserve the private key across restarts and never paste it into logs. Health becomes available only after successful strict model/tokenizer loading. Use another terminal for checks.

   CommandCopy command

   cd "$HOME/kronos-mini-forecast"
   umask 077
   if [ ! -e server.key ]; then openssl rand -hex 32 > server.key; fi
   test -s server.key
   chmod 600 server.key
   OMP_NUM_THREADS=4 MKL_NUM_THREADS=4 HF_HUB_OFFLINE=1 .venv/bin/python reader-forecast-server.py

   Expected result: Native predictor and tokenizer load locally, fixture is written, and the private serial service binds only to loopback port 8993. Startup time and fit are unknown.

Verify the setup

1. 1. Check authenticated readiness

   In a second server terminal use the dedicated virtual environment explicitly. Do not print the key.

   CommandCopy command

   cd "$HOME/kronos-mini-forecast"
   KRONOS_KEY=$(cat server.key)
   curl --fail --silent --show-error -H "Authorization: Bearer $KRONOS_KEY" http://127.0.0.1:8993/health > health.json
   .venv/bin/python - <<'PYHEALTH'
   import json
   x=json.load(open('health.json'))
   assert x['status']=='ready' and x['device']=='cpu' and x['dtype']=='float32'
   assert (x['context_positions'],x['history_positions'],x['forecast_positions'])==(128,64,8)
   print('ready fixed synthetic forecast endpoint')
   PYHEALTH

   Expected result: Ready response reports CPU float32, 128 configured paired-code positions, 64 history and eight forecast positions.

2. 2. Forecast the original fixed synthetic sequence

   Submit exactly one fixture selector. The fixture contains 64 original five-minute OHLCVA rows from a synthetic timezone-naive calendar; it is not an exchange calendar. The request does not accept parameters or arbitrary data. Verify fixture bytes, shape, future timestamps, finite numbers and explicit native units. No exact forecast values or quality threshold are asserted.

   CommandCopy command

   cd "$HOME/kronos-mini-forecast"
   KRONOS_KEY=$(cat server.key)
   curl --fail --silent --show-error --max-time 600 -H "Authorization: Bearer $KRONOS_KEY" -H 'Content-Type: application/json' --data '{"fixture":"synthetic-ohlcva-v1"}' http://127.0.0.1:8993/v1/forecast > forecast.json
   .venv/bin/python - <<'PYCHECK'
   import json,math,datetime,hashlib
   assert hashlib.sha256(open('synthetic-fixture.json','rb').read()).hexdigest()=='f532d0bd540cabb9fabd8107af42986a83f62d5246a025f6bd657c6217f74506'
   x=json.load(open('forecast.json'))
   assert x['columns']==['open','high','low','close','volume','amount']
   assert (x['history_positions'],x['forecast_positions'],x['context_positions'])==(64,8,128)
   assert x['scalar_codes_per_position']==2 and x['generated_scalar_ids_per_path']==16
   assert x['generated_text_tokens'] is None and x['sample_count']==1 and x['seed']==20251016
   start=datetime.datetime(2025,1,2,9,30)
   assert x['timestamps']==[(start+datetime.timedelta(minutes=5*i)).isoformat() for i in range(64,72)]
   assert len(x['forecast'])==8
   assert all(len(r)==6 and all(type(v) in (int,float) and math.isfinite(v) for v in r) for r in x['forecast'])
   print('eight finite six-value forecast rows, exact future timestamps and explicit paired-code units')
   PYCHECK

   Expected result: Eight finite rows with six named values and exact future timestamps pass the checks. A client timeout is not cancellation; stop the foreground process if needed. This expected behavior was not model-executed by the author.

3. 3. Verify the private access boundary

   Health requires the same Bearer key as forecasting. An unauthenticated request must fail.

   CommandCopy command

   cd "$HOME/kronos-mini-forecast"
   test "$(curl --silent --output unauthorized.json --write-out '%{http_code}' http://127.0.0.1:8993/health)" = 401
   .venv/bin/python - <<'PYAUTH'
   import json
   assert json.load(open('unauthorized.json'))=={'error':'unauthorized'}
   print('unauthenticated health request rejected')
   PYAUTH

   Expected result: HTTP 401 and structured unauthorized JSON; no model work is triggered.

Performance evidence

No compatible measurement available. Documented support and estimated memory do not establish response speed.

Limits and troubleshooting

• The custom wrapper is an educational private serial endpoint, not an official or production server. It accepts only the original fixed synthetic fixture; no arbitrary dataset, file path, network URL, market symbol or variable horizon is supported.
• The unit in context, input and output fields is one autoregressive bar position containing two scalar 10-bit IDs. Each stream has 1024 possible IDs. Eight output positions correspond to 16 scalar generated IDs per sampled path and decode to eight rows of six values, not text.
• The configured 128-position window is deliberately below mini published 2048-position context. Sixty-four history plus eight future positions fit without rollover. Native predictor encodes history before window cropping; this wrapper bounds input before calling it and does not silently truncate. Small/base checkpoints and the unreleased large variant remain separate unresolved jobs.
• Both networks explicitly use eval mode and CPU float32. Source Hub safetensors loading does not itself call eval, and native predictor only moves devices. Strict checkpoint loading and saved-config assertions are required. No model execution, peak-memory or latency measurement was performed.
• Native preprocessing computes mean and standard deviation over provided history, normalizes using std plus a small stabilizer, clips normalized values to minus five through five, and inverse-scales predictions. All six columns are supplied to avoid missing-volume/amount defaults. Minute, hour, weekday, day and month are encoded; year, seconds, timezone and exchange calendars are not.
• Native decoded output is unconstrained: prices or volumes may be negative, and OHLC relationships need not hold. Finite shape checks are interface checks, not forecast accuracy, market validity, safety, investment advice or profit guarantees. No real financial data or trading is used.
• Fixed seed aids investigation but is not a guarantee of identical numerical results across runtime versions or hardware. Client timeout does not cancel in-progress CPU work; the server has no generation deadline. Keep it on trusted loopback.

Checkpoint or config mismatch

If strict loading or saved-config assertions fail, stop and verify the exact mini predictor and matching Tokenizer-2k snapshots and source checkout. Do not weaken strict loading or substitute the base tokenizer.

Invalid forecast result

If results contain nonfinite values or the expected eight-by-six shape/index fails, the wrapper returns a generic failure; retain the pinned environment and investigate offline. Do not treat shape success as a quality result.

Insufficient resources

If startup exhausts memory or is too slow, stop the process. Allocated RAM and threads are not measured fit or speed evidence; no reduced-precision or remote fallback is configured.

HTTP 401

If HTTP 401 occurs, read the existing private server.key only in the client shell and send a single Bearer header. Never print, commit or upload the key.

Request timeout

If a request hangs, the 600-second client limit only stops waiting. Use Ctrl+C in the foreground server terminal to stop native work before retrying.

Stop the model

1. 1. Stop the private forecasting process

   Press Ctrl+C in the foreground server terminal. In each client shell unset the key; confirm loopback no longer serves. A process stop does not stop cloud instance billing.

   CommandCopy command

   unset KRONOS_KEY
   if curl --silent --max-time 2 http://127.0.0.1:8993/health >/dev/null; then echo "Port still responds; inspect the foreground process"; exit 1; fi

   Expected result: Foreground process stopped and loopback port no longer responds.

Uninstall

1. 1. Remove dedicated research files

   Preserve desired outputs and license notices first. Confirm the dedicated path before removal. Shared Ubuntu packages remain.

   CommandCopy command

   rm -r -- "$HOME/kronos-mini-forecast"

   Expected result: Dedicated weights, source checkout, virtual environment, wrapper, fixture, results and private key removed.

Shut down cloud resources

1. 1. End cloud billing resources

   In EC2 us-east-1, stop for later reuse or terminate this exact VM. Review Delete on termination for its root volume. Delete unwanted detached EBS volumes, snapshots and reserved/Elastic/public IP resources. Stopped VM storage and retained resources may still bill. Check Bills/Cost Explorer later for final usage.

   Expected result: Exact instance is stopped/terminated as intended; no unwanted attached/retained billed resources remain. Instance compute stops only after its lifecycle transition, not when the Python service stops.