Step 3.5 Flash

Exact model and requirements

Artifact:

StepFun / Full native Step 3.5 Flash BF16 with three MTP layers · checkpoint stepfun-ai/Step-3.5-Flash

Publisher revision:

ab446a3de5e171ea341227e24bb1f090e1b771f7

Artifact digest:

Whole-composition digest not published. All 44 weight parts have exact native LFS SHA256 and sizes; nine auxiliary files use native Git blob SHA1 metadata and sizes in immutable manifest.

Download size:

371.38 GiB

Memory requirement (unknown):

Unknown — No measured per-device or whole-process peak for full BF16/MTP3 on eight H100s at8192 context.

System RAM / disk:

Unknown / Unknown

Placement:

supported sharding. Eight-way tensor parallelism plus expert parallelism; all target/expert/MTP weights retained, no aggregate-fit inference. All target and three MTP layers are sharded on eight GPUs; no CPU or disk weight offload. Host loading buffers and every target/draft cache remain additional and unmeasured.

KV cache:

BF16 attention KV across native full and sliding GQA groups; no Mamba or compressed KV substitution

Workload:

8,192 token context; Unknown; Unknown; concurrency 1; batch 1; cache cold; repetitions Unknown.

Tokenizer:

Exact native Step tokenizer and 5004-byte publisher Jinja · ab446a3de5e171ea341227e24bb1f090e1b771f7

Sampling / reasoning:

Publisher reasoning/agent recommendation temperature 1.0 and top_p 0.95; n=1 and nonstreaming. Raw generated token IDs and native EOS IDs 1/2/128007 are checked. Native template starts assistant thinking. Require raw closing-think 128799 and tool-call 128008/128009 delimiters on first turn, native non-length completion and nonempty final answer. Retain reasoning_content, tool call and inert result exactly; publisher template itself defines which history is rendered. This 8192/4096 demonstration is not the publisher full reasoning budget.

Runtime prerequisites:

Ubuntu 24.04 x86_64 host; pinned Ubuntu 22.04 / Python 3.12 source defaults container; NVIDIA Linux driver >=580.126.20, Working Docker GPU support with NVIDIA Container Toolkit; actual visibility preflight required.

Official artifact · License

Publisher native model card expressly grants Apache License 2.0; full linked license is archived. Review all conditions and applicable third-party rights. Preserve license and relevant notices on redistribution and state changes; patent termination and warranty/liability exclusions apply. No license acceptance or legal clearance is performed here.

• config.json · 0 GiB · revision ab446a3de5e171ea341227e24bb1f090e1b771f7
• tokenizer.json · 0.01 GiB · revision ab446a3de5e171ea341227e24bb1f090e1b771f7
• tokenizer_config.json · 0 GiB · revision ab446a3de5e171ea341227e24bb1f090e1b771f7
• chat_template.jinja · 0 GiB · revision ab446a3de5e171ea341227e24bb1f090e1b771f7
• model.safetensors.index.json · 0 GiB · revision ab446a3de5e171ea341227e24bb1f090e1b771f7
• configuration_step3p5.py · 0 GiB · revision ab446a3de5e171ea341227e24bb1f090e1b771f7
• modeling_step3p5.py · 0 GiB · revision ab446a3de5e171ea341227e24bb1f090e1b771f7
• special_tokens_map.json · 0 GiB · revision ab446a3de5e171ea341227e24bb1f090e1b771f7

• 398,768,720,200 serialized weight bytes are retained in full. Nominal 196B and 11B active do not mean=11B resident weights; metadata reports 199,384,301,376 scalar parameters including auxiliary tensors.
• Each GPU advertises80 GB; exact usable bytes, sharding balance, MTP buffers, attention KV, workspace and host loading overhead are unmeasured. Eight times advertised capacity is not a measured-fit guarantee.

Install the runtime

1. 1. Prepare an existing GPU host and private workspace

   Prepare the already provisioned dedicated eight-H100 Ubuntu 24.04 host with working NVIDIA driver >=580.126.20. Install host utilities and Docker; this guide does not install or claim compatibility of an arbitrary GPU driver. Record actual hardware and stop on any prerequisite failure.

   CommandCopy command

   set -eu
   sudo apt-get update
   sudo apt-get install -y docker.io python3 curl ca-certificates openssl
   sudo systemctl enable --now docker
   mkdir -p "$HOME/step35-flash"
   cd "$HOME/step35-flash"
   nvidia-smi --query-gpu=name,memory.total,driver_version --format=csv,noheader
   python3 - <<'PYPREFLIGHT'
   import csv,subprocess,shutil
   r=list(csv.reader(subprocess.check_output(["nvidia-smi","--query-gpu=name,memory.total,driver_version","--format=csv,noheader,nounits"],text=True).splitlines()))
   assert len(r)==8 and all("H100" in x[0] and "80GB" in x[0].replace(" ","") for x in r)
   assert all(tuple(map(int,x[2].strip().split("."))) >= (580,126,20) for x in r)
   assert shutil.disk_usage(".").free >=600*1024**3
   print("Reader GPU/driver and prescribed600GiB free-disk preflight passed; no model fit measured.")
   PYPREFLIGHT
   sudo docker version

   Expected result: All eight H100 cards and required driver are visible, with prescribed host RAM/disk and free space recorded. Capacity is planning, not measured fit.

2. 2. Configure NVIDIA GPU containers on this dedicated host

   Use the official production repository and exact documented Container Toolkit 1.20.1-1 packages. This modifies Docker daemon configuration and restarts Docker on this dedicated host; preserve its original configuration and do not apply on a shared production daemon without its owner review. Stop on repository/signature/package errors.

   CommandCopy command

   set -euo pipefail
   sudo apt-get install -y gnupg2
   curl --fail --silent --show-error --location https://nvidia.github.io/libnvidia-container/gpgkey | sudo gpg --batch --yes --dearmor -o /usr/share/keyrings/nvidia-container-toolkit-keyring.gpg
   curl --fail --silent --show-error --location https://nvidia.github.io/libnvidia-container/stable/deb/nvidia-container-toolkit.list | sed 's#deb https://#deb [signed-by=/usr/share/keyrings/nvidia-container-toolkit-keyring.gpg] https://#g' | sudo tee /etc/apt/sources.list.d/nvidia-container-toolkit.list >/dev/null
   sudo apt-get update
   sudo apt-get install -y nvidia-container-toolkit=1.20.1-1 nvidia-container-toolkit-base=1.20.1-1 libnvidia-container-tools=1.20.1-1 libnvidia-container1=1.20.1-1
   sudo nvidia-ctk runtime configure --runtime=docker
   sudo systemctl restart docker

   Expected result: The exact GPU container packages install and Docker is configured. The subsequent container nvidia-smi check must pass; this is not model inference or fit proof.

Download the model

1. 1. Write pinned manifest block 1 of 4

   Keep every block in this order. These are exact metadata bytes only; the final block reconstructs the immutable53-entry manifest. No model body is embedded.

   CommandCopy command

   set -eu
   cd "$HOME/step35-flash"
   cat > assets.b64 <<'ASSETBLOCK'
   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
   ASSETBLOCK

   Expected result: Manifest metadata block is written; all blocks are required before the checker can pass.

2. 2. Write pinned manifest block 2 of 4

   Keep every block in this order. These are exact metadata bytes only; the final block reconstructs the immutable53-entry manifest. No model body is embedded.

   CommandCopy command

   set -eu
   cd "$HOME/step35-flash"
   cat >> assets.b64 <<'ASSETBLOCK'
   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
   ASSETBLOCK

   Expected result: Manifest metadata block is written; all blocks are required before the checker can pass.

3. 3. Write pinned manifest block 3 of 4

   Keep every block in this order. These are exact metadata bytes only; the final block reconstructs the immutable53-entry manifest. No model body is embedded.

   CommandCopy command

   set -eu
   cd "$HOME/step35-flash"
   cat >> assets.b64 <<'ASSETBLOCK'
   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
   ASSETBLOCK

   Expected result: Manifest metadata block is written; all blocks are required before the checker can pass.

4. 4. Write pinned manifest block 4 of 4

   Keep every block in this order. These are exact metadata bytes only; the final block reconstructs the immutable53-entry manifest. No model body is embedded.

   CommandCopy command

   set -eu
   cd "$HOME/step35-flash"
   cat >> assets.b64 <<'ASSETBLOCK'
   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
   ASSETBLOCK
   python3 -c "import base64,pathlib; pathlib.Path('assets.json').write_bytes(base64.b64decode(pathlib.Path('assets.b64').read_bytes()))"

   Expected result: Manifest metadata block is written; all blocks are required before the checker can pass.

5. 5. Write streaming integrity helper

   The helper binds the exact manifest before using any entries. Native LFS files use SHA256; native regular Git files use Git blob SHA1, explicitly not a fabricated SHA256. Missing or replaced manifests fail closed.

   CommandCopy command

   set -eu
   cd "$HOME/step35-flash"
   cat > reader-assets.py <<'PYFILE'
   """Reader-run full checkpoint integrity; no model imports."""
   import hashlib
   import json
   import pathlib
   import sys
   import urllib.request
   ROOT = pathlib.Path(__file__).resolve().parent
   MANIFEST_SHA = '2621750ee5e425256187fd012585a2edef6f7eb8308f43214a6abedda3cb55b5'
   
   
   def verified_manifest():
       raw = (ROOT / 'assets.json').read_bytes()
       if hashlib.sha256(raw).hexdigest() != MANIFEST_SHA:
           raise ValueError('manifest_hash')
       d = json.loads(raw)
       if d['revision'] != 'ab446a3de5e171ea341227e24bb1f090e1b771f7' or d['repository'] != 'stepfun-ai/Step-3.5-Flash' or len(d['files']) != 53:
           raise ValueError('manifest_identity')
       return d['files']
   
   
   def check(path, entry):
       digest = hashlib.sha256() if entry['sha256'] else hashlib.sha1()
       if not entry['sha256']:
           digest.update(('blob ' + str(entry['bytes']) + '\0').encode())
       count = 0
       with path.open('rb') as f:
           while block := f.read(1024 * 1024):
               count += len(block)
               digest.update(block)
       if count != entry['bytes'] or digest.hexdigest() != (entry['sha256'] or entry['gitBlobSha1']):
           raise ValueError('asset_integrity_' + entry['path'])
   
   
   def main(download):
       files = verified_manifest()
       out = ROOT / 'model'
       out.mkdir(exist_ok=True)
       for entry in files:
           dest = out / entry['path']
           if pathlib.PurePosixPath(entry['path']).name != entry['path']:
               raise ValueError('asset_path')
           if download and not dest.exists():
               temp = dest.with_suffix(dest.suffix + '.partial')
               with urllib.request.urlopen(entry['url'], timeout=60) as r, temp.open('wb') as f:
                   total = 0
                   while block := r.read(1024 * 1024):
                       total += len(block)
                       if total > entry['bytes']:
                           raise ValueError('download_size')
                       f.write(block)
               check(temp, entry)
               temp.replace(dest)
           check(dest, entry)
       cfg = json.loads((out / 'config.json').read_text())
       if cfg['hidden_size'] != 4096 or cfg['num_hidden_layers'] != 45 or cfg['num_nextn_predict_layers'] != 3 or cfg['moe_num_experts'] != 288 or cfg['moe_top_k'] != 8 or cfg['sliding_window'] != 512:
           raise ValueError('native_config_identity')
       index = json.loads((out / 'model.safetensors.index.json').read_text())['weight_map']
       if set(index.values()) != {e['path'] for e in files if e['path'].endswith('.safetensors')} or not all(any(k.startswith('model.layers.' + str(i) + '.') for k in index) for i in (45,46,47)):
           raise ValueError('full_native_mtp_index')
       print('All53 assets passed exact size and streaming native LFS SHA256 or Git blob SHA1 checks. All44 weight parts and all3 MTP layers remain required.')
   
   
   if __name__ == '__main__':
       if sys.argv[1:] not in (['download'], ['verify']):
           raise SystemExit('Use download or verify')
       main(sys.argv[1] == 'download')
   
   PYFILE

   Expected result: Exact manifest and every complete native part are mandatory.

6. 6. Download and verify all native assets

   Download the exact revision-addressed files. Any access, size, hash, native identity or complete three-MTP index failure stops before server startup. No checkpoint conversion occurs.

   CommandCopy command

   set -eu
   cd "$HOME/step35-flash"
   python3 reader-assets.py download
   python3 reader-assets.py verify

   Expected result: All 53 exact assets pass streaming checks, including all 44 full-weight parts and MTP layers 45,46,47.

Configure

1. 1. Create a private endpoint key

   Create a new private key and environment file in this dedicated reader workspace. Do not disclose it in logs or source control. These are local reader operations, not third-party credentials.

   CommandCopy command

   set -eu
   cd "$HOME/step35-flash"
   umask 077
   python3 - <<'PYPREWRITE'
   import os,pathlib,stat
   paths=[pathlib.Path(n)for n in [".api-key","private.env"]]
   for p in paths:
    try:s=p.lstat()
    except FileNotFoundError:continue
    if not stat.S_ISREG(s.st_mode):raise ValueError('existing secret target must be a regular file; no links')
   for p in paths:
    if p.exists():os.chmod(p,0o600)
   PYPREWRITE
   openssl rand -hex 32 > .api-key
   printf 'VLLM_API_KEY=%s\n' "$(cat .api-key)" > private.env
   chmod 600 .api-key private.env

   Expected result: Private key files have owner-only permissions; no secret is printed.

2. 2. Check the exact released native runtime image

   Use the immutable official Linux amd64 vLLM 0.27.1 image. It contains native Step3p5 target/MTP/config classes; no custom Transformers or remote Python code replacement is needed. Reader pull/version/package and GPU visibility checks are required; source qualification is not execution proof.

   CommandCopy command

   set -eu
   cd "$HOME/step35-flash"
   sudo docker pull --platform linux/amd64 vllm/vllm-openai@sha256:c2f3b1b964e47809b722b5e75b61b1e7b39a50f70388cf2bf2418f16a9f31da2
   sudo docker run --rm --gpus all --entrypoint nvidia-smi vllm/vllm-openai@sha256:c2f3b1b964e47809b722b5e75b61b1e7b39a50f70388cf2bf2418f16a9f31da2 -L
   sudo docker run --rm --entrypoint python3 vllm/vllm-openai@sha256:c2f3b1b964e47809b722b5e75b61b1e7b39a50f70388cf2bf2418f16a9f31da2 -c "import vllm; assert vllm.__version__ == '0.27.1'"
   sudo docker run --rm --entrypoint python3 vllm/vllm-openai@sha256:c2f3b1b964e47809b722b5e75b61b1e7b39a50f70388cf2bf2418f16a9f31da2 -m pip check
   sudo docker run --rm --entrypoint python3 vllm/vllm-openai@sha256:c2f3b1b964e47809b722b5e75b61b1e7b39a50f70388cf2bf2418f16a9f31da2 -m pip freeze > realized-runtime-packages.txt

   Expected result: The exact image/version and dependencies pass; eight GPUs are visible. Any failure stops the walkthrough.

Start the model

1. 1. Start full native BF16 target and three-layer MTP

   Run in a foreground terminal. Released 0.27.1 has its own Step-specific multi-layer proposer; unlike the older card example, retain native num_nextn_predict_layers 3 and select three speculative tokens. Tensor and expert parallelism use all eight GPUs. Do not override the model to one MTP layer or strip weights. The 1800-second demo process bound is not an inference SLA.

   CommandCopy command

   set -eu
   cd "$HOME/step35-flash"
   python3 reader-assets.py verify
   [ "$(nvidia-smi -L | wc -l)" -eq 8 ]
   sudo docker network inspect step35-private >/dev/null 2>&1 || sudo docker network create --internal step35-private
   timeout --signal=TERM --kill-after=10s 1800s sudo docker run --rm --name step35-native --gpus all --shm-size 32g --network step35-private -p 127.0.0.1:8996:8000 --env-file private.env -e HF_HUB_OFFLINE=1 -e TRANSFORMERS_OFFLINE=1 -v "$PWD/model:/model:ro" --entrypoint /bin/bash vllm/vllm-openai@sha256:c2f3b1b964e47809b722b5e75b61b1e7b39a50f70388cf2bf2418f16a9f31da2 -lc 'exec vllm serve /model --served-model-name step35-flash-native --host 0.0.0.0 --port 8000 --dtype bfloat16 --kv-cache-dtype auto --tensor-parallel-size 8 --enable-expert-parallel --disable-cascade-attn --max-model-len 8192 --max-num-seqs 1 --gpu-memory-utilization 0.90 --enforce-eager --no-enable-prefix-caching --reasoning-parser step3p5 --enable-auto-tool-choice --tool-call-parser step3p5 --chat-template /model/chat_template.jinja --speculative-config "{\"method\":\"mtp\",\"num_speculative_tokens\":3}" --generation-config vllm'

   Expected result: Wait for actual readiness before verification. Stop on loader, NCCL, kernel, native parser or OOM errors; no measured startup/fit claim.

Verify the setup

1. 1. Write native protocol verifier block 1 of 2

   Keep both blocks in order. The exact corrected client rejects residual or malformed native control output and requires closed thinking, native EOS and integer token counts. No model operation was performed by the author.

   CommandCopy command

   set -eu
   cd "$HOME/step35-flash"
   cat > reader-client.py <<'CLIENTBLOCK'
   """Fixed inert tool/reasoning check for exact native Step3.5 Flash; stdlib only."""
   import http.client
   import json
   import pathlib
   
   MODEL = 'step35-flash-native'
   CONTEXT = 8192
   OUTPUT = 4096
   KWARGS = {}  # Exact publisher template controls native thinking and history.
   TOOLS = [{'type': 'function', 'function': {'name': 'lookup_demo_constant',
       'description': 'Look up the inert demonstration constant; no external action.',
       'parameters': {'type': 'object', 'properties': {'key': {'type': 'string',
       'enum': ['alpha']}}, 'required': ['key'], 'additionalProperties': False}}}]
   
   
   def unique(pairs):
       out = {}
       for k, v in pairs:
           if k in out:
               raise ValueError('duplicate_json_key')
           out[k] = v
       return out
   
   
   def parse(raw):
       if type(raw) is not bytes or len(raw) > 2 * 1024 * 1024:
           raise ValueError('response_bytes')
       return json.loads(raw.decode('utf-8'), object_pairs_hook=unique,
           parse_constant=lambda _: (_ for _ in ()).throw(ValueError('nonfinite_json')))
   
   
   def budget(d):
       if type(d) is not dict or type(d.get('count')) is not int or not 1 <= d['count'] <= CONTEXT - OUTPUT:
           raise ValueError('prompt_budget')
       if d.get('max_model_len') != CONTEXT:
           raise ValueError('wrong_server_context')
       if type(d.get('max_model_len')) is not int:
           raise ValueError('server_context_type')
       return d['count']
   
   
   def checked(d, prompt, tool_turn):
       if type(d) is not dict or d.get('model') != MODEL or type(d.get('choices')) is not list or len(d['choices']) != 1:
           raise ValueError('response_identity')
       if type(prompt) is not int or not 1 <= prompt <= CONTEXT - OUTPUT:
           raise ValueError('prompt_count_type')
       choice = d['choices'][0]
       if type(choice) is not dict:
           raise ValueError('choice_shape')
       if choice.get('stop_reason') is not None and type(choice.get('stop_reason')) is not int:
           raise ValueError('stop_reason_type')
       expected = 'tool_calls' if tool_turn else 'stop'
       if choice.get('finish_reason') != expected or choice.get('stop_reason') not in (None, 1, 2, 128007):
           raise ValueError('not_native_complete')
       ids = choice.get('token_ids')
       if type(ids) is not list or not 1 <= len(ids) < OUTPUT or any(type(i) is not int or i < 0 for i in ids):
           raise ValueError('raw_output_tokens')
       # The native prompt already opens thinking. Generated output may echo one
       # opening token, but must close exactly once before the answer/tool segment.
       if ids.count(128799) != 1 or ids.count(128798) > 1:
           raise ValueError('native_thinking_boundary')
       close = ids.index(128799)
       if 128798 in ids and ids.index(128798) >= close:
           raise ValueError('reopened_thinking')
       eos = [i for i, token in enumerate(ids) if token in (1,2,128007)]
       # Released scheduler appends the sampled stop token before stopping; the
       # detokenizer retains it in raw IDs even when omitting its decoded text.
       if len(eos) != 1 or eos[0] != len(ids)-1:
           raise ValueError('nonterminal_native_eos')
       if choice.get('stop_reason') is not None and choice['stop_reason'] != ids[-1]:
           raise ValueError('native_stop_identity')
       prefix = ids[:close]
       if any(token >= 128000 and token != 128798 for token in prefix):
           raise ValueError('control_inside_reasoning')
       body = ids[close+1:]
       if tool_turn:
           if body.count(128008) != 1 or body.count(128009) != 1 or body.index(128008) >= body.index(128009):
               raise ValueError('native_tool_boundary')
           if body[0] != 128008 or body[-2] != 128009:
               raise ValueError('native_tool_only_body')
           allowed = {128008,128009,128007}
       else:
           allowed = {128007}
       if any(token >= 128000 and token not in allowed for token in body):
           raise ValueError('unexpected_native_control')
       usage = d.get('usage', {})
       if type(usage) is not dict or any(type(usage.get(k)) is not int for k in ('prompt_tokens','completion_tokens','total_tokens')):
           raise ValueError('usage_integer_types')
       if usage.get('prompt_tokens') != prompt or usage.get('completion_tokens') != len(ids) or usage.get('total_tokens') != prompt + len(ids):
           raise ValueError('usage_mismatch')
       if prompt + len(ids) > CONTEXT:
           raise ValueError('total_budget')
       message = choice.get('message')
       if type(message) is not dict or message.get('role') != 'assistant' or type(message.get('reasoning')) is not str or not message['reasoning'].strip():
           raise ValueError('native_reasoning_required')
       markers = ('<tool', '</tool', '<function', '</function', '<parameter',
                  '</parameter', '<think', '</think', '<|', '<｜', '<im_', '<patch_')
       if any(x in message['reasoning'] for x in markers):
           raise ValueError('parsed_reasoning_control')
       content = message.get('content')
       if content is not None and type(content) is not str:
           raise ValueError('content_type')
       if content and any(x in content for x in markers):
           raise ValueError('parsed_content_control')
       if tool_turn:
           if content and content.strip():
               raise ValueError('unexpected_tool_turn_content')
           calls = message.get('tool_calls')
           if type(calls) is not list or len(calls) != 1 or 128008 not in ids or 128009 not in ids:
               raise ValueError('complete_one_native_tool')
           call = calls[0]
           if call.get('type') != 'function' or type(call.get('id')) is not str or not 1 <= len(call['id']) <= 128:
               raise ValueError('tool_identity')
           fn = call.get('function', {})
           if fn.get('name') != 'lookup_demo_constant' or type(fn.get('arguments')) is not str:
               raise ValueError('tool_name')
   CLIENTBLOCK

   Expected result: The complete exact helper is reconstructed and checksum-verified; it never executes tools or changes returned content.

2. 2. Write native protocol verifier block 2 of 2

   Keep both blocks in order. The exact corrected client rejects residual or malformed native control output and requires closed thinking, native EOS and integer token counts. No model operation was performed by the author.

   CommandCopy command

   set -eu
   cd "$HOME/step35-flash"
   cat >> reader-client.py <<'CLIENTBLOCK'
           args = parse(fn['arguments'].encode())
           if type(args) is not dict or args != {'key': 'alpha'}:
               raise ValueError('fixed_tool_arguments')
       elif message.get('tool_calls') or type(message.get('content')) is not str or not message['content'].strip():
           raise ValueError('final_answer_required')
       return message
   
   
   def call(path, body, key):
       payload = json.dumps(body, allow_nan=False).encode()
       if len(payload) > 65536:
           raise ValueError('fixed_client_body_bound')
       c = http.client.HTTPConnection('127.0.0.1', 8996, timeout=120)
       try:
           c.request('POST', path, payload, {'Authorization': 'Bearer ' + key,
                                           'Content-Type': 'application/json'})
           r = c.getresponse()
           if r.status != 200:
               raise RuntimeError('endpoint_status_' + str(r.status))
           return parse(r.read(2 * 1024 * 1024 + 1))
       finally:
           c.close()
   
   
   def run(request):
       messages = [{'role': 'user', 'content': 'Use lookup_demo_constant for key alpha, then explain in one sentence why its value plus one is eight. Treat this only as synthetic protocol testing.'}]
       outputs = []
       for tool_turn in (True, False):
           shared = {'model': MODEL, 'messages': messages, 'tools': TOOLS,
                     'chat_template_kwargs': KWARGS, 'add_special_tokens': False,
                     'add_generation_prompt': True}
           count = budget(request('/tokenize', shared))
           response = request('/v1/chat/completions', {**shared, 'tool_choice': 'auto' if tool_turn else 'none',
               'temperature': 1.0, 'top_p': 0.95, 'max_tokens': OUTPUT, 'n': 1,
               'stream': False, 'return_token_ids': True, 'stop_token_ids': [1, 2, 128007],
               'include_reasoning': True})
           msg = checked(response, count, tool_turn)
           outputs.append(response)
           if tool_turn:
               # Preserve complete native reasoning, final body and tool call history. No executor.
               messages.append({'role': 'assistant', 'content': msg.get('content') or '',
                   'reasoning_content': msg['reasoning'], 'tool_calls': msg['tool_calls']})
               messages.append({'role': 'tool', 'tool_call_id': msg['tool_calls'][0]['id'],
                                'content': '{"key":"alpha","value":7,"synthetic":true}'})
       return outputs
   
   
   if __name__ == '__main__':
       key = (pathlib.Path(__file__).resolve().parent / '.api-key').read_text().strip()
       if len(key) != 64 or any(c not in '0123456789abcdef' for c in key):
           raise SystemExit('private_key_format')
       result = run(lambda path, body: call(path, body, key))
       pathlib.Path('protocol-result.json').write_text(json.dumps(result, indent=2, allow_nan=False) + '\n')
       print('Native raw closing-think/tool boundaries and token budgets passed. Inspect final answer and record semantic failures; this is not a quality or performance benchmark.')
   CLIENTBLOCK
   python3 -m py_compile reader-client.py
   sha256sum --check <<'CHECKSUM'
   d86637c87bf3b2e68a87a58b703c76eff32bb72cf6ea31b194b6f3febc0c5f56  reader-client.py
   CHECKSUM

   Expected result: The complete exact helper is reconstructed and checksum-verified; it never executes tools or changes returned content.

3. 3. Check readiness, auth boundary and native completion

   Native health/tokenize routes are not Bearer-protected; the server remains bound to host loopback. Unauthenticated /v1/models must return 401. No public port is opened. The client has a 120-second socket timeout that does not cancel GPU work.

   CommandCopy command

   set -eu
   cd "$HOME/step35-flash"
   curl --fail --silent http://127.0.0.1:8996/health
   [ "$(curl --silent -o /dev/null -w '%{http_code}' http://127.0.0.1:8996/v1/models)" = 401 ]
   python3 reader-client.py

   Expected result: Record actual protocol-result.json and human semantic assessment. Residual tool/role/thinking controls, repeated or unclosed thinking, malformed tool order, nonterminal or missing native EOS, noninteger usage, overbudget or length finish fail closed. Protocol validity is not task quality.

Performance evidence

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

Limits and troubleshooting

• Native 45 target layers and three MTP layers are preserved; all 288 routed experts plus shared expert are retained. Only 8 routed experts activate per token, which does not reduce full resident checkpoint requirements.
• Native full attention alternates with 512-window sliding attention. Config max_seq_len 262144 and tokenizer model_max_length 131072 differ; this 8192-position walkthrough establishes neither maximum nor fit.
• Released vLLM 0.27.1 Step3p5MTP uses its own multi-group/per-step proposer and all three prediction layers. This source-qualified choice supersedes the older card one-layer override; no speculative acceptance, speed or quality has been measured.
• Native reasoning and XML tool delimiters differ from generic JSON-only tool formatting. The exact publisher Jinja remains unchanged; only parsed inert lookup arguments are accepted by the client, and no tool executes.
• The fixed client 120-second socket timeout does not cancel native work; 1800-second foreground lifetime is a chosen demonstration bound. Stop/kill the exact container when needed.
• Weights, target/draft activations, per-GPU caches and NCCL/workspace overhead remain unmeasured. Advertised 8 × 80 GB capacity and 2048 GiB host allocation do not prove fit.
• Original repository/date/provider alias is preserved; exact native checkpoint pin is not asserted byte-equal to any provider API. Broader FP8, long-context, alternate hardware/runtime and task scopes remain open.

Native loading, collective initialization or GPU out-of-memory fails

Stop the exact container. Recheck all assets, eight GPUs, driver/image versions and source-qualified flags. Record failure; do not strip MTP/experts, silently quantize or change the checkpoint.

Missing closing thinking/tool markers, wrong parsed arguments or length finish

Treat the run as failed. Keep the exact request/history/token budget and retain diagnostics. Do not execute model-generated calls or infer task correctness from valid JSON alone.

Stop the model

1. 1. Stop the exact private container

   Stop the foreground process and remove only this exact named service container. Client timeout does not guarantee server cancellation.

   CommandCopy command

   sudo docker stop --time 10 step35-native || true
   if [ "$(sudo docker inspect -f '{{.State.Running}}' step35-native 2>/dev/null || true)" = true ]; then sudo docker kill step35-native; fi
   curl --max-time 3 http://127.0.0.1:8996/v1/models || true

   Expected result: The exact service container and loopback endpoint stop. Owned-host power/storage costs may remain.

Uninstall

1. 1. Remove dedicated reader assets and network

   Remove this dedicated reader directory, internal network and exact image only after preserving desired results. Do not prune shared Docker resources or uninstall shared GPU drivers/toolkit.

   CommandCopy command

   sudo docker rm -f step35-native 2>/dev/null || true
   sudo docker network rm step35-private
   sudo docker image rm vllm/vllm-openai@sha256:c2f3b1b964e47809b722b5e75b61b1e7b39a50f70388cf2bf2418f16a9f31da2
   rm -r -- "$HOME/step35-flash"

   Expected result: Dedicated reader artifacts are removed; shared Docker, drivers and unrelated resources remain.

Exact model and requirements

Artifact:

StepFun / Full native Step 3.5 Flash BF16 with three MTP layers · checkpoint stepfun-ai/Step-3.5-Flash

Publisher revision:

ab446a3de5e171ea341227e24bb1f090e1b771f7

Artifact digest:

Whole-composition digest not published. All 44 weight parts have exact native LFS SHA256 and sizes; nine auxiliary files use native Git blob SHA1 metadata and sizes in immutable manifest.

Download size:

371.38 GiB

Memory requirement (unknown):

Unknown — No measured per-device or whole-process peak for full BF16/MTP3 on eight H100s at8192 context.

System RAM / disk:

Unknown / Unknown

Placement:

supported sharding. Eight-way tensor parallelism plus expert parallelism; all target/expert/MTP weights retained, no aggregate-fit inference. All target and three MTP layers are sharded on eight GPUs; no CPU or disk weight offload. Host loading buffers and every target/draft cache remain additional and unmeasured.

KV cache:

BF16 attention KV across native full and sliding GQA groups; no Mamba or compressed KV substitution

Workload:

8,192 token context; Unknown; Unknown; concurrency 1; batch 1; cache cold; repetitions Unknown.

Tokenizer:

Exact native Step tokenizer and 5004-byte publisher Jinja · ab446a3de5e171ea341227e24bb1f090e1b771f7

Sampling / reasoning:

Publisher reasoning/agent recommendation temperature 1.0 and top_p 0.95; n=1 and nonstreaming. Raw generated token IDs and native EOS IDs 1/2/128007 are checked. Native template starts assistant thinking. Require raw closing-think 128799 and tool-call 128008/128009 delimiters on first turn, native non-length completion and nonempty final answer. Retain reasoning_content, tool call and inert result exactly; publisher template itself defines which history is rendered. This 8192/4096 demonstration is not the publisher full reasoning budget.

Runtime prerequisites:

Ubuntu 24.04 x86_64 host; pinned Ubuntu 22.04 / Python 3.12 source defaults container; NVIDIA Linux driver >=580.126.20, Working Docker GPU support with NVIDIA Container Toolkit; actual visibility preflight required.

Official artifact · License

Publisher native model card expressly grants Apache License 2.0; full linked license is archived. Review all conditions and applicable third-party rights. Preserve license and relevant notices on redistribution and state changes; patent termination and warranty/liability exclusions apply. No license acceptance or legal clearance is performed here.

• config.json · 0 GiB · revision ab446a3de5e171ea341227e24bb1f090e1b771f7
• tokenizer.json · 0.01 GiB · revision ab446a3de5e171ea341227e24bb1f090e1b771f7
• tokenizer_config.json · 0 GiB · revision ab446a3de5e171ea341227e24bb1f090e1b771f7
• chat_template.jinja · 0 GiB · revision ab446a3de5e171ea341227e24bb1f090e1b771f7
• model.safetensors.index.json · 0 GiB · revision ab446a3de5e171ea341227e24bb1f090e1b771f7
• configuration_step3p5.py · 0 GiB · revision ab446a3de5e171ea341227e24bb1f090e1b771f7
• modeling_step3p5.py · 0 GiB · revision ab446a3de5e171ea341227e24bb1f090e1b771f7
• special_tokens_map.json · 0 GiB · revision ab446a3de5e171ea341227e24bb1f090e1b771f7

• 398,768,720,200 serialized weight bytes are retained in full. Nominal 196B and 11B active do not mean=11B resident weights; metadata reports 199,384,301,376 scalar parameters including auxiliary tensors.
• Each GPU advertises80 GB; exact usable bytes, sharding balance, MTP buffers, attention KV, workspace and host loading overhead are unmeasured. Eight times advertised capacity is not a measured-fit guarantee.

Install the runtime

1. 1. Prepare an existing GPU host and private workspace

   Prepare the already provisioned dedicated eight-H100 Ubuntu 24.04 host with working NVIDIA driver >=580.126.20. Install host utilities and Docker; this guide does not install or claim compatibility of an arbitrary GPU driver. Record actual hardware and stop on any prerequisite failure.

   CommandCopy command

   set -eu
   sudo apt-get update
   sudo apt-get install -y docker.io python3 curl ca-certificates openssl
   sudo systemctl enable --now docker
   mkdir -p "$HOME/step35-flash"
   cd "$HOME/step35-flash"
   nvidia-smi --query-gpu=name,memory.total,driver_version --format=csv,noheader
   python3 - <<'PYPREFLIGHT'
   import csv,subprocess,shutil
   r=list(csv.reader(subprocess.check_output(["nvidia-smi","--query-gpu=name,memory.total,driver_version","--format=csv,noheader,nounits"],text=True).splitlines()))
   assert len(r)==8 and all("H100" in x[0] and "80GB" in x[0].replace(" ","") for x in r)
   assert all(tuple(map(int,x[2].strip().split("."))) >= (580,126,20) for x in r)
   assert shutil.disk_usage(".").free >=600*1024**3
   print("Reader GPU/driver and prescribed600GiB free-disk preflight passed; no model fit measured.")
   PYPREFLIGHT
   sudo docker version

   Expected result: All eight H100 cards and required driver are visible, with prescribed host RAM/disk and free space recorded. Capacity is planning, not measured fit.

2. 2. Configure NVIDIA GPU containers on this dedicated host

   Use the official production repository and exact documented Container Toolkit 1.20.1-1 packages. This modifies Docker daemon configuration and restarts Docker on this dedicated host; preserve its original configuration and do not apply on a shared production daemon without its owner review. Stop on repository/signature/package errors.

   CommandCopy command

   set -euo pipefail
   sudo apt-get install -y gnupg2
   curl --fail --silent --show-error --location https://nvidia.github.io/libnvidia-container/gpgkey | sudo gpg --batch --yes --dearmor -o /usr/share/keyrings/nvidia-container-toolkit-keyring.gpg
   curl --fail --silent --show-error --location https://nvidia.github.io/libnvidia-container/stable/deb/nvidia-container-toolkit.list | sed 's#deb https://#deb [signed-by=/usr/share/keyrings/nvidia-container-toolkit-keyring.gpg] https://#g' | sudo tee /etc/apt/sources.list.d/nvidia-container-toolkit.list >/dev/null
   sudo apt-get update
   sudo apt-get install -y nvidia-container-toolkit=1.20.1-1 nvidia-container-toolkit-base=1.20.1-1 libnvidia-container-tools=1.20.1-1 libnvidia-container1=1.20.1-1
   sudo nvidia-ctk runtime configure --runtime=docker
   sudo systemctl restart docker

   Expected result: The exact GPU container packages install and Docker is configured. The subsequent container nvidia-smi check must pass; this is not model inference or fit proof.

Download the model

1. 1. Write pinned manifest block 1 of 4

   Keep every block in this order. These are exact metadata bytes only; the final block reconstructs the immutable53-entry manifest. No model body is embedded.

   CommandCopy command

   set -eu
   cd "$HOME/step35-flash"
   cat > assets.b64 <<'ASSETBLOCK'
   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
   ASSETBLOCK

   Expected result: Manifest metadata block is written; all blocks are required before the checker can pass.

2. 2. Write pinned manifest block 2 of 4

   Keep every block in this order. These are exact metadata bytes only; the final block reconstructs the immutable53-entry manifest. No model body is embedded.

   CommandCopy command

   set -eu
   cd "$HOME/step35-flash"
   cat >> assets.b64 <<'ASSETBLOCK'
   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
   ASSETBLOCK

   Expected result: Manifest metadata block is written; all blocks are required before the checker can pass.

3. 3. Write pinned manifest block 3 of 4

   Keep every block in this order. These are exact metadata bytes only; the final block reconstructs the immutable53-entry manifest. No model body is embedded.

   CommandCopy command

   set -eu
   cd "$HOME/step35-flash"
   cat >> assets.b64 <<'ASSETBLOCK'
   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
   ASSETBLOCK

   Expected result: Manifest metadata block is written; all blocks are required before the checker can pass.

4. 4. Write pinned manifest block 4 of 4

   Keep every block in this order. These are exact metadata bytes only; the final block reconstructs the immutable53-entry manifest. No model body is embedded.

   CommandCopy command

   set -eu
   cd "$HOME/step35-flash"
   cat >> assets.b64 <<'ASSETBLOCK'
   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
   ASSETBLOCK
   python3 -c "import base64,pathlib; pathlib.Path('assets.json').write_bytes(base64.b64decode(pathlib.Path('assets.b64').read_bytes()))"

   Expected result: Manifest metadata block is written; all blocks are required before the checker can pass.

5. 5. Write streaming integrity helper

   The helper binds the exact manifest before using any entries. Native LFS files use SHA256; native regular Git files use Git blob SHA1, explicitly not a fabricated SHA256. Missing or replaced manifests fail closed.

   CommandCopy command

   set -eu
   cd "$HOME/step35-flash"
   cat > reader-assets.py <<'PYFILE'
   """Reader-run full checkpoint integrity; no model imports."""
   import hashlib
   import json
   import pathlib
   import sys
   import urllib.request
   ROOT = pathlib.Path(__file__).resolve().parent
   MANIFEST_SHA = '2621750ee5e425256187fd012585a2edef6f7eb8308f43214a6abedda3cb55b5'
   
   
   def verified_manifest():
       raw = (ROOT / 'assets.json').read_bytes()
       if hashlib.sha256(raw).hexdigest() != MANIFEST_SHA:
           raise ValueError('manifest_hash')
       d = json.loads(raw)
       if d['revision'] != 'ab446a3de5e171ea341227e24bb1f090e1b771f7' or d['repository'] != 'stepfun-ai/Step-3.5-Flash' or len(d['files']) != 53:
           raise ValueError('manifest_identity')
       return d['files']
   
   
   def check(path, entry):
       digest = hashlib.sha256() if entry['sha256'] else hashlib.sha1()
       if not entry['sha256']:
           digest.update(('blob ' + str(entry['bytes']) + '\0').encode())
       count = 0
       with path.open('rb') as f:
           while block := f.read(1024 * 1024):
               count += len(block)
               digest.update(block)
       if count != entry['bytes'] or digest.hexdigest() != (entry['sha256'] or entry['gitBlobSha1']):
           raise ValueError('asset_integrity_' + entry['path'])
   
   
   def main(download):
       files = verified_manifest()
       out = ROOT / 'model'
       out.mkdir(exist_ok=True)
       for entry in files:
           dest = out / entry['path']
           if pathlib.PurePosixPath(entry['path']).name != entry['path']:
               raise ValueError('asset_path')
           if download and not dest.exists():
               temp = dest.with_suffix(dest.suffix + '.partial')
               with urllib.request.urlopen(entry['url'], timeout=60) as r, temp.open('wb') as f:
                   total = 0
                   while block := r.read(1024 * 1024):
                       total += len(block)
                       if total > entry['bytes']:
                           raise ValueError('download_size')
                       f.write(block)
               check(temp, entry)
               temp.replace(dest)
           check(dest, entry)
       cfg = json.loads((out / 'config.json').read_text())
       if cfg['hidden_size'] != 4096 or cfg['num_hidden_layers'] != 45 or cfg['num_nextn_predict_layers'] != 3 or cfg['moe_num_experts'] != 288 or cfg['moe_top_k'] != 8 or cfg['sliding_window'] != 512:
           raise ValueError('native_config_identity')
       index = json.loads((out / 'model.safetensors.index.json').read_text())['weight_map']
       if set(index.values()) != {e['path'] for e in files if e['path'].endswith('.safetensors')} or not all(any(k.startswith('model.layers.' + str(i) + '.') for k in index) for i in (45,46,47)):
           raise ValueError('full_native_mtp_index')
       print('All53 assets passed exact size and streaming native LFS SHA256 or Git blob SHA1 checks. All44 weight parts and all3 MTP layers remain required.')
   
   
   if __name__ == '__main__':
       if sys.argv[1:] not in (['download'], ['verify']):
           raise SystemExit('Use download or verify')
       main(sys.argv[1] == 'download')
   
   PYFILE

   Expected result: Exact manifest and every complete native part are mandatory.

6. 6. Download and verify all native assets

   Download the exact revision-addressed files. Any access, size, hash, native identity or complete three-MTP index failure stops before server startup. No checkpoint conversion occurs.

   CommandCopy command

   set -eu
   cd "$HOME/step35-flash"
   python3 reader-assets.py download
   python3 reader-assets.py verify

   Expected result: All 53 exact assets pass streaming checks, including all 44 full-weight parts and MTP layers 45,46,47.

Configure

1. 1. Create a private endpoint key

   Create a new private key and environment file in this dedicated reader workspace. Do not disclose it in logs or source control. These are local reader operations, not third-party credentials.

   CommandCopy command

   set -eu
   cd "$HOME/step35-flash"
   umask 077
   python3 - <<'PYPREWRITE'
   import os,pathlib,stat
   paths=[pathlib.Path(n)for n in [".api-key","private.env"]]
   for p in paths:
    try:s=p.lstat()
    except FileNotFoundError:continue
    if not stat.S_ISREG(s.st_mode):raise ValueError('existing secret target must be a regular file; no links')
   for p in paths:
    if p.exists():os.chmod(p,0o600)
   PYPREWRITE
   openssl rand -hex 32 > .api-key
   printf 'VLLM_API_KEY=%s\n' "$(cat .api-key)" > private.env
   chmod 600 .api-key private.env

   Expected result: Private key files have owner-only permissions; no secret is printed.

2. 2. Check the exact released native runtime image

   Use the immutable official Linux amd64 vLLM 0.27.1 image. It contains native Step3p5 target/MTP/config classes; no custom Transformers or remote Python code replacement is needed. Reader pull/version/package and GPU visibility checks are required; source qualification is not execution proof.

   CommandCopy command

   set -eu
   cd "$HOME/step35-flash"
   sudo docker pull --platform linux/amd64 vllm/vllm-openai@sha256:c2f3b1b964e47809b722b5e75b61b1e7b39a50f70388cf2bf2418f16a9f31da2
   sudo docker run --rm --gpus all --entrypoint nvidia-smi vllm/vllm-openai@sha256:c2f3b1b964e47809b722b5e75b61b1e7b39a50f70388cf2bf2418f16a9f31da2 -L
   sudo docker run --rm --entrypoint python3 vllm/vllm-openai@sha256:c2f3b1b964e47809b722b5e75b61b1e7b39a50f70388cf2bf2418f16a9f31da2 -c "import vllm; assert vllm.__version__ == '0.27.1'"
   sudo docker run --rm --entrypoint python3 vllm/vllm-openai@sha256:c2f3b1b964e47809b722b5e75b61b1e7b39a50f70388cf2bf2418f16a9f31da2 -m pip check
   sudo docker run --rm --entrypoint python3 vllm/vllm-openai@sha256:c2f3b1b964e47809b722b5e75b61b1e7b39a50f70388cf2bf2418f16a9f31da2 -m pip freeze > realized-runtime-packages.txt

   Expected result: The exact image/version and dependencies pass; eight GPUs are visible. Any failure stops the walkthrough.

Start the model

1. 1. Start full native BF16 target and three-layer MTP

   Run in a foreground terminal. Released 0.27.1 has its own Step-specific multi-layer proposer; unlike the older card example, retain native num_nextn_predict_layers 3 and select three speculative tokens. Tensor and expert parallelism use all eight GPUs. Do not override the model to one MTP layer or strip weights. The 1800-second demo process bound is not an inference SLA.

   CommandCopy command

   set -eu
   cd "$HOME/step35-flash"
   python3 reader-assets.py verify
   [ "$(nvidia-smi -L | wc -l)" -eq 8 ]
   sudo docker network inspect step35-private >/dev/null 2>&1 || sudo docker network create --internal step35-private
   timeout --signal=TERM --kill-after=10s 1800s sudo docker run --rm --name step35-native --gpus all --shm-size 32g --network step35-private -p 127.0.0.1:8996:8000 --env-file private.env -e HF_HUB_OFFLINE=1 -e TRANSFORMERS_OFFLINE=1 -v "$PWD/model:/model:ro" --entrypoint /bin/bash vllm/vllm-openai@sha256:c2f3b1b964e47809b722b5e75b61b1e7b39a50f70388cf2bf2418f16a9f31da2 -lc 'exec vllm serve /model --served-model-name step35-flash-native --host 0.0.0.0 --port 8000 --dtype bfloat16 --kv-cache-dtype auto --tensor-parallel-size 8 --enable-expert-parallel --disable-cascade-attn --max-model-len 8192 --max-num-seqs 1 --gpu-memory-utilization 0.90 --enforce-eager --no-enable-prefix-caching --reasoning-parser step3p5 --enable-auto-tool-choice --tool-call-parser step3p5 --chat-template /model/chat_template.jinja --speculative-config "{\"method\":\"mtp\",\"num_speculative_tokens\":3}" --generation-config vllm'

   Expected result: Wait for actual readiness before verification. Stop on loader, NCCL, kernel, native parser or OOM errors; no measured startup/fit claim.

Verify the setup

1. 1. Write native protocol verifier block 1 of 2

   Keep both blocks in order. The exact corrected client rejects residual or malformed native control output and requires closed thinking, native EOS and integer token counts. No model operation was performed by the author.

   CommandCopy command

   set -eu
   cd "$HOME/step35-flash"
   cat > reader-client.py <<'CLIENTBLOCK'
   """Fixed inert tool/reasoning check for exact native Step3.5 Flash; stdlib only."""
   import http.client
   import json
   import pathlib
   
   MODEL = 'step35-flash-native'
   CONTEXT = 8192
   OUTPUT = 4096
   KWARGS = {}  # Exact publisher template controls native thinking and history.
   TOOLS = [{'type': 'function', 'function': {'name': 'lookup_demo_constant',
       'description': 'Look up the inert demonstration constant; no external action.',
       'parameters': {'type': 'object', 'properties': {'key': {'type': 'string',
       'enum': ['alpha']}}, 'required': ['key'], 'additionalProperties': False}}}]
   
   
   def unique(pairs):
       out = {}
       for k, v in pairs:
           if k in out:
               raise ValueError('duplicate_json_key')
           out[k] = v
       return out
   
   
   def parse(raw):
       if type(raw) is not bytes or len(raw) > 2 * 1024 * 1024:
           raise ValueError('response_bytes')
       return json.loads(raw.decode('utf-8'), object_pairs_hook=unique,
           parse_constant=lambda _: (_ for _ in ()).throw(ValueError('nonfinite_json')))
   
   
   def budget(d):
       if type(d) is not dict or type(d.get('count')) is not int or not 1 <= d['count'] <= CONTEXT - OUTPUT:
           raise ValueError('prompt_budget')
       if d.get('max_model_len') != CONTEXT:
           raise ValueError('wrong_server_context')
       if type(d.get('max_model_len')) is not int:
           raise ValueError('server_context_type')
       return d['count']
   
   
   def checked(d, prompt, tool_turn):
       if type(d) is not dict or d.get('model') != MODEL or type(d.get('choices')) is not list or len(d['choices']) != 1:
           raise ValueError('response_identity')
       if type(prompt) is not int or not 1 <= prompt <= CONTEXT - OUTPUT:
           raise ValueError('prompt_count_type')
       choice = d['choices'][0]
       if type(choice) is not dict:
           raise ValueError('choice_shape')
       if choice.get('stop_reason') is not None and type(choice.get('stop_reason')) is not int:
           raise ValueError('stop_reason_type')
       expected = 'tool_calls' if tool_turn else 'stop'
       if choice.get('finish_reason') != expected or choice.get('stop_reason') not in (None, 1, 2, 128007):
           raise ValueError('not_native_complete')
       ids = choice.get('token_ids')
       if type(ids) is not list or not 1 <= len(ids) < OUTPUT or any(type(i) is not int or i < 0 for i in ids):
           raise ValueError('raw_output_tokens')
       # The native prompt already opens thinking. Generated output may echo one
       # opening token, but must close exactly once before the answer/tool segment.
       if ids.count(128799) != 1 or ids.count(128798) > 1:
           raise ValueError('native_thinking_boundary')
       close = ids.index(128799)
       if 128798 in ids and ids.index(128798) >= close:
           raise ValueError('reopened_thinking')
       eos = [i for i, token in enumerate(ids) if token in (1,2,128007)]
       # Released scheduler appends the sampled stop token before stopping; the
       # detokenizer retains it in raw IDs even when omitting its decoded text.
       if len(eos) != 1 or eos[0] != len(ids)-1:
           raise ValueError('nonterminal_native_eos')
       if choice.get('stop_reason') is not None and choice['stop_reason'] != ids[-1]:
           raise ValueError('native_stop_identity')
       prefix = ids[:close]
       if any(token >= 128000 and token != 128798 for token in prefix):
           raise ValueError('control_inside_reasoning')
       body = ids[close+1:]
       if tool_turn:
           if body.count(128008) != 1 or body.count(128009) != 1 or body.index(128008) >= body.index(128009):
               raise ValueError('native_tool_boundary')
           if body[0] != 128008 or body[-2] != 128009:
               raise ValueError('native_tool_only_body')
           allowed = {128008,128009,128007}
       else:
           allowed = {128007}
       if any(token >= 128000 and token not in allowed for token in body):
           raise ValueError('unexpected_native_control')
       usage = d.get('usage', {})
       if type(usage) is not dict or any(type(usage.get(k)) is not int for k in ('prompt_tokens','completion_tokens','total_tokens')):
           raise ValueError('usage_integer_types')
       if usage.get('prompt_tokens') != prompt or usage.get('completion_tokens') != len(ids) or usage.get('total_tokens') != prompt + len(ids):
           raise ValueError('usage_mismatch')
       if prompt + len(ids) > CONTEXT:
           raise ValueError('total_budget')
       message = choice.get('message')
       if type(message) is not dict or message.get('role') != 'assistant' or type(message.get('reasoning')) is not str or not message['reasoning'].strip():
           raise ValueError('native_reasoning_required')
       markers = ('<tool', '</tool', '<function', '</function', '<parameter',
                  '</parameter', '<think', '</think', '<|', '<｜', '<im_', '<patch_')
       if any(x in message['reasoning'] for x in markers):
           raise ValueError('parsed_reasoning_control')
       content = message.get('content')
       if content is not None and type(content) is not str:
           raise ValueError('content_type')
       if content and any(x in content for x in markers):
           raise ValueError('parsed_content_control')
       if tool_turn:
           if content and content.strip():
               raise ValueError('unexpected_tool_turn_content')
           calls = message.get('tool_calls')
           if type(calls) is not list or len(calls) != 1 or 128008 not in ids or 128009 not in ids:
               raise ValueError('complete_one_native_tool')
           call = calls[0]
           if call.get('type') != 'function' or type(call.get('id')) is not str or not 1 <= len(call['id']) <= 128:
               raise ValueError('tool_identity')
           fn = call.get('function', {})
           if fn.get('name') != 'lookup_demo_constant' or type(fn.get('arguments')) is not str:
               raise ValueError('tool_name')
   CLIENTBLOCK

   Expected result: The complete exact helper is reconstructed and checksum-verified; it never executes tools or changes returned content.

2. 2. Write native protocol verifier block 2 of 2

   Keep both blocks in order. The exact corrected client rejects residual or malformed native control output and requires closed thinking, native EOS and integer token counts. No model operation was performed by the author.

   CommandCopy command

   set -eu
   cd "$HOME/step35-flash"
   cat >> reader-client.py <<'CLIENTBLOCK'
           args = parse(fn['arguments'].encode())
           if type(args) is not dict or args != {'key': 'alpha'}:
               raise ValueError('fixed_tool_arguments')
       elif message.get('tool_calls') or type(message.get('content')) is not str or not message['content'].strip():
           raise ValueError('final_answer_required')
       return message
   
   
   def call(path, body, key):
       payload = json.dumps(body, allow_nan=False).encode()
       if len(payload) > 65536:
           raise ValueError('fixed_client_body_bound')
       c = http.client.HTTPConnection('127.0.0.1', 8996, timeout=120)
       try:
           c.request('POST', path, payload, {'Authorization': 'Bearer ' + key,
                                           'Content-Type': 'application/json'})
           r = c.getresponse()
           if r.status != 200:
               raise RuntimeError('endpoint_status_' + str(r.status))
           return parse(r.read(2 * 1024 * 1024 + 1))
       finally:
           c.close()
   
   
   def run(request):
       messages = [{'role': 'user', 'content': 'Use lookup_demo_constant for key alpha, then explain in one sentence why its value plus one is eight. Treat this only as synthetic protocol testing.'}]
       outputs = []
       for tool_turn in (True, False):
           shared = {'model': MODEL, 'messages': messages, 'tools': TOOLS,
                     'chat_template_kwargs': KWARGS, 'add_special_tokens': False,
                     'add_generation_prompt': True}
           count = budget(request('/tokenize', shared))
           response = request('/v1/chat/completions', {**shared, 'tool_choice': 'auto' if tool_turn else 'none',
               'temperature': 1.0, 'top_p': 0.95, 'max_tokens': OUTPUT, 'n': 1,
               'stream': False, 'return_token_ids': True, 'stop_token_ids': [1, 2, 128007],
               'include_reasoning': True})
           msg = checked(response, count, tool_turn)
           outputs.append(response)
           if tool_turn:
               # Preserve complete native reasoning, final body and tool call history. No executor.
               messages.append({'role': 'assistant', 'content': msg.get('content') or '',
                   'reasoning_content': msg['reasoning'], 'tool_calls': msg['tool_calls']})
               messages.append({'role': 'tool', 'tool_call_id': msg['tool_calls'][0]['id'],
                                'content': '{"key":"alpha","value":7,"synthetic":true}'})
       return outputs
   
   
   if __name__ == '__main__':
       key = (pathlib.Path(__file__).resolve().parent / '.api-key').read_text().strip()
       if len(key) != 64 or any(c not in '0123456789abcdef' for c in key):
           raise SystemExit('private_key_format')
       result = run(lambda path, body: call(path, body, key))
       pathlib.Path('protocol-result.json').write_text(json.dumps(result, indent=2, allow_nan=False) + '\n')
       print('Native raw closing-think/tool boundaries and token budgets passed. Inspect final answer and record semantic failures; this is not a quality or performance benchmark.')
   CLIENTBLOCK
   python3 -m py_compile reader-client.py
   sha256sum --check <<'CHECKSUM'
   d86637c87bf3b2e68a87a58b703c76eff32bb72cf6ea31b194b6f3febc0c5f56  reader-client.py
   CHECKSUM

   Expected result: The complete exact helper is reconstructed and checksum-verified; it never executes tools or changes returned content.

3. 3. Check readiness, auth boundary and native completion

   Native health/tokenize routes are not Bearer-protected; the server remains bound to host loopback. Unauthenticated /v1/models must return 401. No public port is opened. The client has a 120-second socket timeout that does not cancel GPU work.

   CommandCopy command

   set -eu
   cd "$HOME/step35-flash"
   curl --fail --silent http://127.0.0.1:8996/health
   [ "$(curl --silent -o /dev/null -w '%{http_code}' http://127.0.0.1:8996/v1/models)" = 401 ]
   python3 reader-client.py

   Expected result: Record actual protocol-result.json and human semantic assessment. Residual tool/role/thinking controls, repeated or unclosed thinking, malformed tool order, nonterminal or missing native EOS, noninteger usage, overbudget or length finish fail closed. Protocol validity is not task quality.

Performance evidence

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

Limits and troubleshooting

• Native 45 target layers and three MTP layers are preserved; all 288 routed experts plus shared expert are retained. Only 8 routed experts activate per token, which does not reduce full resident checkpoint requirements.
• Native full attention alternates with 512-window sliding attention. Config max_seq_len 262144 and tokenizer model_max_length 131072 differ; this 8192-position walkthrough establishes neither maximum nor fit.
• Released vLLM 0.27.1 Step3p5MTP uses its own multi-group/per-step proposer and all three prediction layers. This source-qualified choice supersedes the older card one-layer override; no speculative acceptance, speed or quality has been measured.
• Native reasoning and XML tool delimiters differ from generic JSON-only tool formatting. The exact publisher Jinja remains unchanged; only parsed inert lookup arguments are accepted by the client, and no tool executes.
• The fixed client 120-second socket timeout does not cancel native work; 1800-second foreground lifetime is a chosen demonstration bound. Stop/kill the exact container when needed.
• Weights, target/draft activations, per-GPU caches and NCCL/workspace overhead remain unmeasured. Advertised 8 × 80 GB capacity and 2048 GiB host allocation do not prove fit.
• Original repository/date/provider alias is preserved; exact native checkpoint pin is not asserted byte-equal to any provider API. Broader FP8, long-context, alternate hardware/runtime and task scopes remain open.

Native loading, collective initialization or GPU out-of-memory fails

Stop the exact container. Recheck all assets, eight GPUs, driver/image versions and source-qualified flags. Record failure; do not strip MTP/experts, silently quantize or change the checkpoint.

Missing closing thinking/tool markers, wrong parsed arguments or length finish

Treat the run as failed. Keep the exact request/history/token budget and retain diagnostics. Do not execute model-generated calls or infer task correctness from valid JSON alone.

Stop the model

1. 1. Stop the exact private container

   Stop the foreground process and remove only this exact named service container. Client timeout does not guarantee server cancellation.

   CommandCopy command

   sudo docker stop --time 10 step35-native || true
   if [ "$(sudo docker inspect -f '{{.State.Running}}' step35-native 2>/dev/null || true)" = true ]; then sudo docker kill step35-native; fi
   curl --max-time 3 http://127.0.0.1:8996/v1/models || true

   Expected result: The exact service container and loopback endpoint stop. Owned-host power/storage costs may remain.

Uninstall

1. 1. Remove dedicated reader assets and network

   Remove this dedicated reader directory, internal network and exact image only after preserving desired results. Do not prune shared Docker resources or uninstall shared GPU drivers/toolkit.

   CommandCopy command

   sudo docker rm -f step35-native 2>/dev/null || true
   sudo docker network rm step35-private
   sudo docker image rm vllm/vllm-openai@sha256:c2f3b1b964e47809b722b5e75b61b1e7b39a50f70388cf2bf2418f16a9f31da2
   rm -r -- "$HOME/step35-flash"

   Expected result: Dedicated reader artifacts are removed; shared Docker, drivers and unrelated resources remain.

Shut down cloud resources

1. 1. End all instance and retained-resource billing

   In EC2 us-east-1 stop for later reuse or terminate this exact p5.48xlarge. Review Delete on termination for its 768 GiB root volume; remove unwanted detached EBS volumes, snapshots and retained IP resources. Local NVMe is ephemeral. Inspect Bills/Cost Explorer for final use; no cloud actions were performed here.

   Expected result: Compute ends only after the actual instance transition. Retained volumes, snapshots and IP resources can still bill; container stop alone ends no VM bill.