Maple-Preview

Exact model and requirements

Artifact:

DeepGrove AI / Maple Preview publisher packed ternary MLX, dense decoding · checkpoint deepgrove/maple-preview-2bit-mlx

Publisher revision:

e8be321fb325c40beb7b621bd98e8e57a324ec42

Artifact digest:

No composite whole-package digest is published. Every required file has exact size and LFS SHA-256 or native Git blob SHA-1 in the embedded mandatory manifest.

Download size:

4.95 GiB

Memory requirement (unknown):

Unknown — No complete native runtime, cache, activation, loading-buffer or OS peak was measured.

System RAM / disk:

Unknown / Unknown

Placement:

single device. One selected accelerator; all layers and experts remain present. Complete model and caches use the integrated GPU/unified pool, with no disk or CPU weight offload.

KV cache:

Native MLX cache dtype follows packed model activations; exact realized dtype/peak is not measured

Workload:

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

Tokenizer:

Exact native Maple tokenizer and 3,292-byte publisher Jinja, with no added BOS · e8be321fb325c40beb7b621bd98e8e57a324ec42

Sampling / reasoning:

Dense greedy argmax is an author-selected fixed-task setting. Native packed dense head is selected explicitly; no FlashHead approximation or performance transfer. The publisher template opens thinking. Preserve complete reasoning_content and final answer in the second turn. Require exactly closed thinking and true raw native EOS; reject partial/reopened/tool/control output. 8,192 total context and 4,096 output cap are walkthrough bounds, not the native full-context recommendation.

Runtime prerequisites:

macOS 26.2 or newer, arm64, Python 3.12; Apple Metal GPU and exact MLX native ABI.

Official artifact · License

Exact publisher MIT license is archived. Retain its copyright and permission notice in copies or substantial portions; software is provided without warranty or liability. The separate publisher MLX fork is also MIT. No terms are accepted or legal clearance granted by this guide.

• README.md · 0 GiB · revision e8be321fb325c40beb7b621bd98e8e57a324ec42
• added_tokens.json · 0 GiB · revision e8be321fb325c40beb7b621bd98e8e57a324ec42
• chat_template.jinja · 0 GiB · revision e8be321fb325c40beb7b621bd98e8e57a324ec42
• config.json · 0 GiB · revision e8be321fb325c40beb7b621bd98e8e57a324ec42
• maple.py · 0 GiB · revision e8be321fb325c40beb7b621bd98e8e57a324ec42
• merges.txt · 0 GiB · revision e8be321fb325c40beb7b621bd98e8e57a324ec42
• model.safetensors.index.json · 0 GiB · revision e8be321fb325c40beb7b621bd98e8e57a324ec42
• special_tokens_map.json · 0 GiB · revision e8be321fb325c40beb7b621bd98e8e57a324ec42
• tokenizer.json · 0.01 GiB · revision e8be321fb325c40beb7b621bd98e8e57a324ec42
• tokenizer_config.json · 0 GiB · revision e8be321fb325c40beb7b621bd98e8e57a324ec42
• vocab.json · 0 GiB · revision e8be321fb325c40beb7b621bd98e8e57a324ec42

• 5314328134 serialized weight bytes do not include complete runtime memory.
• All experts, caches, native extension workspaces, activations, loading buffers and OS overhead remain additional and unmeasured.

Install the runtime

1. 1. Prepare the existing reader machine

   Use the exact platform in prerequisites. Record actual capacity before downloading. These are planning checks, not measured model fit.

   CommandCopy command

   set -eu
   mkdir -p "$HOME/maple-preview-local-m4-32gb-mlx-dense"
   cd "$HOME/maple-preview-local-m4-32gb-mlx-dense"
   test "$(uname -m)" = arm64
   sw_vers
   sysctl -n hw.memsize
   command -v python3.12
   python3.12 -m venv .venv
   python3.12 - <<'PYPREFLIGHT'
   import shutil
   assert shutil.disk_usage('.').free >= 26843545600
   print('Reader prescribed free space passed; fit remains unmeasured.')
   PYPREFLIGHT

   Expected result: An existing compatible machine, Python 3.12 environment and prescribed free space are recorded. Stop on platform, GPU, driver or disk failure.

Download the model

1. 1. Write the mandatory streaming asset checker (1/1)

   Write this exact source section in order. Writing source does not execute native code.

   CommandCopy command

   set -eu
   cd "$HOME/maple-preview-local-m4-32gb-mlx-dense"
   cat > reader_assets.py <<'MAPLEFILE'
   """Reader-only exact full Maple asset verification; no native imports."""
   import hashlib,json,pathlib,sys,urllib.request
   ROOT=pathlib.Path(__file__).resolve().parent
   PINS={'cuda':('deepgrove/maple-preview','ac1ddd79d2b5cb4406f5d2bebdf95406ce505a07','3ab6ce508a608d0af564c6683497eb7d166178d8c992e33dd41404c9e2be92eb',23),'mlx':('deepgrove/maple-preview-2bit-mlx','e8be321fb325c40beb7b621bd98e8e57a324ec42','d236df54f11ada042e099c2ae15942b246c611420495ee39ac3a470a38b5a182',15)}
   def manifest(mode):
       repo,pin,digest,count=PINS[mode];data=(ROOT/'assets.json').read_bytes()
       if hashlib.sha256(data).hexdigest()!=digest:raise ValueError('manifest_hash')
       m=json.loads(data)
       if m['mode']!=mode or m['repo']!=repo or m['revision']!=pin or len(m['files'])!=count:raise ValueError('manifest_identity')
       return m
   
   def check(path,row):
       h=hashlib.sha256()if row['algorithm']=='sha256'else hashlib.sha1();n=0
       if row['algorithm']=='git-sha1':h.update(b'blob '+str(row['bytes']).encode()+b'\0')
       with path.open('rb')as f:
           for b in iter(lambda:f.read(1024*1024),b''):n+=len(b);h.update(b)
       if n!=row['bytes']or h.hexdigest()!=row['digest']:raise ValueError('asset_integrity_'+row['path'])
   
   def verify(mode,download=False):
       m=manifest(mode)
       for category,dirname in [('files','model'),('runtime_files','runtime')]:
           root=ROOT/dirname;root.mkdir(exist_ok=True)
           for row in m[category]:
               if pathlib.PurePosixPath(row['path']).name!=row['path']:raise ValueError('asset_path')
               path=root/row['path']
               if download and not path.exists():
                   temp=path.with_suffix(path.suffix+'.partial');n=0
                   with urllib.request.urlopen(row['url'],timeout=120)as src,temp.open('wb')as dst:
                       for chunk in iter(lambda:src.read(1024*1024),b''):
                           n+=len(chunk)
                           if n>row['bytes']:raise ValueError('download_size')
                           dst.write(chunk)
                   check(temp,row);temp.replace(path)
               check(path,row)
       index=json.loads((ROOT/'model/model.safetensors.index.json').read_text())['weight_map'];parts={x['path']for x in m['files']if x['path'].endswith('.safetensors')}
       # Publisher MLX index covers the three main parts; FlashHead is a mandatory
       # separately shipped fourth asset, loaded by publisher model*.safetensors glob.
       indexed=parts-({'model-flashhead.safetensors'}if mode=='mlx'else set())
       if set(index.values())!=indexed:raise ValueError('complete_weight_index')
       print('All exact native or publisher-packed files verified; no fit or numerical equivalence proof.')
   if __name__=='__main__':
       if len(sys.argv)!=3 or sys.argv[1]not in PINS or sys.argv[2]not in ('download','verify'):raise SystemExit('Use cuda|mlx download|verify')
       verify(sys.argv[1],sys.argv[2]=='download')
   MAPLEFILE

   Expected result: The exact source section is written. Continue through every section before verification.

2. 2. Write the exact complete publisher asset manifest (1/1)

   Write this exact source section in order. Writing source does not execute native code.

   CommandCopy command

   set -eu
   cd "$HOME/maple-preview-local-m4-32gb-mlx-dense"
   cat > assets.json <<'MAPLEFILE'
   {"mode":"mlx","repo":"deepgrove/maple-preview-2bit-mlx","revision":"e8be321fb325c40beb7b621bd98e8e57a324ec42","files":[{"path":"README.md","bytes":107,"algorithm":"git-sha1","digest":"8cec12c472d936a3b1e0043e2cabe97450156788","url":"https://huggingface.co/deepgrove/maple-preview-2bit-mlx/resolve/e8be321fb325c40beb7b621bd98e8e57a324ec42/README.md"},{"path":"added_tokens.json","bytes":707,"algorithm":"git-sha1","digest":"b54f9135e44c1e81047e8d05cb027af8bc039eed","url":"https://huggingface.co/deepgrove/maple-preview-2bit-mlx/resolve/e8be321fb325c40beb7b621bd98e8e57a324ec42/added_tokens.json"},{"path":"chat_template.jinja","bytes":3292,"algorithm":"git-sha1","digest":"b5509b8aef363d7262fb86370db1eb41f554959b","url":"https://huggingface.co/deepgrove/maple-preview-2bit-mlx/resolve/e8be321fb325c40beb7b621bd98e8e57a324ec42/chat_template.jinja"},{"path":"config.json","bytes":2710,"algorithm":"git-sha1","digest":"f71e7d2c090c3da042828e4f6b6fc292f7e04cb8","url":"https://huggingface.co/deepgrove/maple-preview-2bit-mlx/resolve/e8be321fb325c40beb7b621bd98e8e57a324ec42/config.json"},{"path":"maple.py","bytes":71017,"algorithm":"git-sha1","digest":"7190efe661bef2df8aff90ebd42962e37129f421","url":"https://huggingface.co/deepgrove/maple-preview-2bit-mlx/resolve/e8be321fb325c40beb7b621bd98e8e57a324ec42/maple.py"},{"path":"merges.txt","bytes":1671853,"algorithm":"git-sha1","digest":"31349551d90c7606f325fe0f11bbb8bd5fa0d7c7","url":"https://huggingface.co/deepgrove/maple-preview-2bit-mlx/resolve/e8be321fb325c40beb7b621bd98e8e57a324ec42/merges.txt"},{"path":"model-00001-of-00003.safetensors","bytes":2162084350,"algorithm":"sha256","digest":"1444bbc92e2842e9a612c32b6d6f30f0f026a107c8094a4ac273819a22eb84e1","url":"https://huggingface.co/deepgrove/maple-preview-2bit-mlx/resolve/e8be321fb325c40beb7b621bd98e8e57a324ec42/model-00001-of-00003.safetensors"},{"path":"model-00002-of-00003.safetensors","bytes":2187444586,"algorithm":"sha256","digest":"d051e252c46570f989b65c5e037b94402d22f892e96dc91f5fb83dbb06b50d30","url":"https://huggingface.co/deepgrove/maple-preview-2bit-mlx/resolve/e8be321fb325c40beb7b621bd98e8e57a324ec42/model-00002-of-00003.safetensors"},{"path":"model-00003-of-00003.safetensors","bytes":958711742,"algorithm":"sha256","digest":"85c97321214c864597562d5b7286b9e8a0dac36e5abd622f11f9b8425ee88705","url":"https://huggingface.co/deepgrove/maple-preview-2bit-mlx/resolve/e8be321fb325c40beb7b621bd98e8e57a324ec42/model-00003-of-00003.safetensors"},{"path":"model-flashhead.safetensors","bytes":6087456,"algorithm":"sha256","digest":"67d1af06cd8ec6f434ea57b9e066dedf1f7572a95d1af14c56a525cb8b7b03cd","url":"https://huggingface.co/deepgrove/maple-preview-2bit-mlx/resolve/e8be321fb325c40beb7b621bd98e8e57a324ec42/model-flashhead.safetensors"},{"path":"model.safetensors.index.json","bytes":40054,"algorithm":"git-sha1","digest":"6bbc2e03c0de04a8feca573322479e1a3ea78738","url":"https://huggingface.co/deepgrove/maple-preview-2bit-mlx/resolve/e8be321fb325c40beb7b621bd98e8e57a324ec42/model.safetensors.index.json"},{"path":"special_tokens_map.json","bytes":613,"algorithm":"git-sha1","digest":"ac23c0aaa2434523c494330aeb79c58395378103","url":"https://huggingface.co/deepgrove/maple-preview-2bit-mlx/resolve/e8be321fb325c40beb7b621bd98e8e57a324ec42/special_tokens_map.json"},{"path":"tokenizer.json","bytes":11422654,"algorithm":"sha256","digest":"aeb13307a71acd8fe81861d94ad54ab689df773318809eed3cbe794b4492dae4","url":"https://huggingface.co/deepgrove/maple-preview-2bit-mlx/resolve/e8be321fb325c40beb7b621bd98e8e57a324ec42/tokenizer.json"},{"path":"tokenizer_config.json","bytes":5432,"algorithm":"git-sha1","digest":"f4b91854389fd13e964c436e576a0011e588104b","url":"https://huggingface.co/deepgrove/maple-preview-2bit-mlx/resolve/e8be321fb325c40beb7b621bd98e8e57a324ec42/tokenizer_config.json"},{"path":"vocab.json","bytes":2776833,"algorithm":"git-sha1","digest":"4783fe10ac3adce15ac8f358ef5462739852c569","url":"https://huggingface.co/deepgrove/maple-preview-2bit-mlx/resolve/e8be321fb325c40beb7b621bd98e8e57a324ec42/vocab.json"}],"runtime_files":[{"path":"mlx_lm_maple_kernels-0.1.0-cp312-cp312-macosx_26_0_arm64.whl","bytes":49175,"algorithm":"git-sha1","digest":"a6cd4f72c2ef5b8d6d8a901061c4ad5c8b1e39cc","url":"https://huggingface.co/deepgrove/maple-preview-2bit-mlx/resolve/ad3e2a0a772a2131271483ba6770a7cbb9846b6f/wheels/mlx_lm_maple_kernels-0.1.0-cp312-cp312-macosx_26_0_arm64.whl"}]}
   MAPLEFILE

   Expected result: The exact source section is written. Continue through every section before verification.

3. 3. Verify the manifest and download every required asset

   Only the reader runs this download. All model parts, tokenizer auxiliaries, remote code and the Apple kernel wheel are mandatory. Existing mismatched files fail closed; no byte range or alternate checkpoint is used.

   CommandCopy command

   set -eu
   cd "$HOME/maple-preview-local-m4-32gb-mlx-dense"
   .venv/bin/python - <<'PYHASH'
   import hashlib,pathlib
   for n,h in [('reader_assets.py', 'e12f5b1a39737ba1649b235702ebe76404406d7422e1593c3a20b20a34fe56fe'), ('assets.json', 'd236df54f11ada042e099c2ae15942b246c611420495ee39ac3a470a38b5a182')]:
    assert hashlib.sha256(pathlib.Path(n).read_bytes()).hexdigest()==h
   PYHASH
   .venv/bin/python reader_assets.py mlx download

   Expected result: All required per-file sizes and streaming hashes pass before any native loading. This establishes exact bytes, not model quality or fit.

Configure

1. 1. Install the exact native-facing reader runtime

   Reader installation is source-qualified and unexecuted here. Use a fresh environment; no FA3, Liger, alternate MLX fork or floating kernel build. Stop on wheel ABI, dependency or native import failure. Save actual transitive versions; this is not an executed lockfile.

   CommandCopy command

   set -eu
   cd "$HOME/maple-preview-local-m4-32gb-mlx-dense"
   git clone https://github.com/deepgrove-ai/mlx-lm-deepgrove.git runtime-src
   git -C runtime-src checkout --detach 151f62d1b811bd0af7c5365ab0fbaa57e3a64113
   test "$(git -C runtime-src rev-parse HEAD)" = 151f62d1b811bd0af7c5365ab0fbaa57e3a64113
   .venv/bin/python -m pip install "https://files.pythonhosted.org/packages/0f/01/02110ceacf4efd00ec172f60b4cd42c8b1509ac50ffa65a6a77d723133aa/mlx-0.32.0-cp312-cp312-macosx_26_0_arm64.whl#sha256=8dfb577faa4dc413cfd0d6eb78f230d3b3b6169df4473e84408abdeb21346e9d" transformers==5.11.0 numpy==2.2.3 sentencepiece==0.2.1 protobuf==6.32.1 PyYAML==6.0.2 Jinja2==3.1.6
   .venv/bin/python reader_assets.py mlx verify
   .venv/bin/python -m pip install --no-deps "runtime/mlx_lm_maple_kernels-0.1.0-cp312-cp312-macosx_26_0_arm64.whl"
   .venv/bin/python -m pip install --no-deps ./runtime-src
   .venv/bin/python -m pip check
   .venv/bin/python -m pip freeze > realized-environment.txt

   Expected result: Exact native-facing versions install and pip check passes only after reader execution. Record realized versions; no compatibility or fit was measured by the author.

2. 2. Write the complete private native reasoning service (1/3)

   Write this exact source section in order. Writing source does not execute native code.

   CommandCopy command

   set -eu
   cd "$HOME/maple-preview-local-m4-32gb-mlx-dense"
   cat > maple_server.py <<'MAPLEFILE'
   """Reader-authored fixed Maple reasoning/code harness. Native imports occur only in reader startup."""
   import hashlib,hmac,json,pathlib,sys
   from http.server import BaseHTTPRequestHandler,HTTPServer
   ROOT=pathlib.Path(__file__).resolve().parent
   CONTEXT=8192;CAP=4096;EOS=151645;OPEN=151667;CLOSE=151668
   PROMPTS=['A library opens at 09:00 and closes at 17:00 on the same day. Work out its opening duration, then give a small Python function opening_hours() returning that integer. Do not execute code or call tools.','Now closing changes to 18:00 with the same opening time. Recompute and update the function. Do not execute code or call tools.']
   def native_forward(model,input_ids,cache):
       # Exact native no-mask, no-explicit-position path. Never model.generate().
       return model(input_ids=input_ids,attention_mask=None,position_ids=None,
           past_key_values=cache,use_cache=True,cache_position=None,return_dict=True,logits_to_keep=1)
   
   def parse_output(ids,text):
       if type(ids)is not list or not 1<=len(ids)<=CAP or any(type(x)is not int or x<0 for x in ids):raise ValueError('raw_output_ids')
       if ids[-1]!=EOS or ids.count(EOS)!=1 or ids.count(CLOSE)!=1 or OPEN in ids:raise ValueError('native_end_or_think')
       end=ids.index(CLOSE)
       if any(x>=151643 and x not in (OPEN,CLOSE,EOS)for x in ids):raise ValueError('unexpected_native_control_id')
       if type(text)is not str or text.count('</think>')!=1:raise ValueError('decoded_thinking_boundary')
       reasoning,answer=text.split('</think>')
       if not reasoning.strip()or not answer.strip()or any(x in reasoning+answer for x in ('<think','</think','<tool','</tool','<|','|>')):raise ValueError('nonfinal_control_or_empty')
       return reasoning,answer
   
   def checked_result(mode,d):
       if type(d)is not dict or set(d)!={'backend','turns','context_limit','output_cap','head'}or d['backend']!=mode or d['context_limit']!=CONTEXT or d['output_cap']!=CAP or d['head']!='dense':raise ValueError('result_identity')
       if type(d['turns'])is not list or len(d['turns'])!=2:raise ValueError('two_turns_required')
       for r in d['turns']:
           if type(r)is not dict or set(r)!={'reasoning','answer','input_tokens','output_tokens','stop_id'}:raise ValueError('turn_schema')
           if any(type(r[k])is not int for k in ('input_tokens','output_tokens','stop_id'))or not 1<=r['input_tokens']<=CONTEXT-CAP or not 1<=r['output_tokens']<=CAP or r['input_tokens']+r['output_tokens']>CONTEXT or r['stop_id']!=EOS:raise ValueError('turn_budget_stop')
           if any(type(r[k])is not str or not r[k].strip() or len(r[k])>65536 for k in ('reasoning','answer')):raise ValueError('turn_text')
           if any(c in r['reasoning']+r['answer']for c in ('<think','</think','<tool','</tool','<|','|>')):raise ValueError('turn_control')
       return d
   
   def publisher_native_extension(model):
       # Publisher loader executes custom_model without registering sys.modules.
       # The verified source-defined class initializer retains that module globals.
       init=model.__class__.__init__
       context=getattr(init,'__globals__',None)
       if type(context)is not dict or context.get('__name__')!=model.__class__.__module__:
           raise RuntimeError('publisher_loaded_context_required')
       native=context.get('_maple_native')
       if native is None:raise RuntimeError('pinned_publisher_kernel_required')
       return native
   
   def make_engine(mode):
       from reader_assets import verify
       verify(mode)
       path=ROOT/'model';cfg=json.loads((path/'config.json').read_text())
       if cfg['num_hidden_layers']!=24 or cfg['num_experts']!=256 or cfg['num_experts_per_tok']!=8 or cfg['sliding_window']!=512:raise RuntimeError('wrong_maple_architecture')
       if mode=='cuda':
           import torch
           from transformers import AutoTokenizer,AutoModelForCausalLM
           if torch.__version__.split('+')[0]!='2.6.0'or torch.version.cuda!='12.4'or torch.cuda.device_count()!=1 or 'H100'not in torch.cuda.get_device_name(0):raise RuntimeError('exact_cuda_runtime_required')
           # Select only the explicit FA2 fallback; FA3 is deliberately not installed.
           import importlib.util
           if importlib.util.find_spec('flash_attn_interface')is not None:raise RuntimeError('unexpected_fa3_interface')
           import flash_attn
           if flash_attn.__version__!='2.8.3':raise RuntimeError('flash_attention_version')
           import inspect
           if 'window_size'not in inspect.signature(flash_attn.flash_attn_func).parameters or importlib.util.find_spec('liger_kernel')is not None:raise RuntimeError('exact_unfused_sliding_attention_backend')
           if cfg['max_position_embeddings']!=131072 or cfg['nope_on_global_attention']is not True:raise RuntimeError('native_cuda_config')
           tok=AutoTokenizer.from_pretrained(path,local_files_only=True,trust_remote_code=False)
   MAPLEFILE

   Expected result: The exact source section is written. Continue through every section before verification.

3. 3. Write the complete private native reasoning service (2/3)

   Write this exact source section in order. Writing source does not execute native code.

   CommandCopy command

   set -eu
   cd "$HOME/maple-preview-local-m4-32gb-mlx-dense"
   cat >> maple_server.py <<'MAPLEFILE'
           model,info=AutoModelForCausalLM.from_pretrained(path,local_files_only=True,trust_remote_code=True,torch_dtype=torch.bfloat16,device_map=None,use_safetensors=True,output_loading_info=True,ignore_mismatched_sizes=False)
           keys={'missing_keys','unexpected_keys','mismatched_keys','error_msgs'}
           if type(info)is not dict or set(info)!=keys or any(info[k]for k in keys):raise RuntimeError('native_load_diagnostics')
           model=model.to('cuda:0').eval()
           if len(model.model.layers)!=24 or any(p.dtype!=torch.bfloat16 or p.device.type!='cuda' or not torch.isfinite(p).all()for p in model.parameters()):raise RuntimeError('full_native_gpu_parameters')
           def generate(ids):
               tokens=[];cache=None;current=torch.tensor([ids],dtype=torch.long,device='cuda:0')
               with torch.inference_mode():
                   for _ in range(CAP):
                       out=native_forward(model,current,cache);cache=out.past_key_values;scores=out.logits[0,-1]
                       if not torch.isfinite(scores).all():raise RuntimeError('nonfinite_native_logits')
                       token=int(torch.argmax(scores));tokens.append(token)
                       if token==EOS:break
                       current=torch.tensor([[token]],dtype=torch.long,device='cuda:0')
               return tokens,tok.decode(tokens[:-1]if tokens and tokens[-1]==EOS else tokens,skip_special_tokens=False)
       elif mode=='mlx':
           import platform,subprocess
           if platform.machine()!='arm64'or tuple(map(int,platform.mac_ver()[0].split('.')[:2]))<(26,2):raise RuntimeError('native_apple_platform')
           import mlx.core as mx
           from mlx_lm import load
           from mlx_lm.generate import stream_generate
           if mx.__version__!='0.32.0'or mx.default_device()!=mx.gpu:raise RuntimeError('exact_mlx_metal_runtime')
           if cfg['max_position_embeddings']!=128000 or cfg.get('nope_on_global_attention')is not False:raise RuntimeError('publisher_mlx_config')
           model,tok=load(str(path),model_config={'use_flash_head':False},trust_remote_code=True)
           if model.lm_head_flash is not None or len(model.model.layers)!=24:raise RuntimeError('dense_complete_maple_required')
           import importlib.metadata
           native=publisher_native_extension(model)
           if importlib.metadata.version('mlx-lm-maple-kernels')!='0.1.0'or native is None:raise RuntimeError('pinned_publisher_kernel_required')
           def sampler(logprobs):
               if bool(mx.any(mx.isnan(logprobs)))or bool(mx.any(mx.isinf(logprobs)&(logprobs>0)))or not bool(mx.any(mx.isfinite(logprobs))):raise RuntimeError('nonfinite_dense_logits')
               return mx.argmax(logprobs,axis=-1)
           def generate(ids):
               tokens=[];last=None
               for item in stream_generate(model,tok,ids,max_tokens=CAP,sampler=sampler):
                   tokens.append(item.token);last=item
               if last is None or last.finish_reason!='stop' or last.generation_tokens!=len(tokens):raise RuntimeError('native_mlx_completion')
               return tokens,tok.decode(tokens[:-1]if tokens and tokens[-1]==EOS else tokens,skip_special_tokens=False)
       else:raise ValueError('backend')
       if tok.eos_token_id!=EOS or tok.convert_tokens_to_ids('</think>')!=CLOSE:raise RuntimeError('native_token_identity')
       def engine():
           messages=[];turns=[]
           for prompt in PROMPTS:
               messages.append({'role':'user','content':prompt})
               rendered=tok.apply_chat_template(messages,tokenize=False,add_generation_prompt=True)
               ids=tok.encode(rendered,add_special_tokens=False)
               if not 1<=len(ids)<=CONTEXT-CAP:raise ValueError('full_native_history_budget')
               raw,text=generate(ids);reasoning,answer=parse_output(raw,text)
               turns.append({'reasoning':reasoning,'answer':answer,'input_tokens':len(ids),'output_tokens':len(raw),'stop_id':raw[-1]})
               messages.append({'role':'assistant','content':answer,'reasoning_content':reasoning})
           return checked_result(mode,{'backend':mode,'turns':turns,'context_limit':CONTEXT,'output_cap':CAP,'head':'dense'})
       return engine
   
   def create_server(engine,key,mode,port=9008):
       if type(key)is not str or len(key)!=64 or any(x not in '0123456789abcdef'for x in key):raise ValueError('key_format')
       class Handler(BaseHTTPRequestHandler):
           def setup(self):super().setup();self.connection.settimeout(5)
           def log_message(self,*_):pass
           def reply(self,status,d):
               data=json.dumps(d,allow_nan=False).encode()
               if len(data)>131072:status=500;data=b'{"error":"response_limit"}'
   MAPLEFILE

   Expected result: The exact source section is written. Continue through every section before verification.

4. 4. Write the complete private native reasoning service (3/3)

   Write this exact source section in order. Writing source does not execute native code.

   CommandCopy command

   set -eu
   cd "$HOME/maple-preview-local-m4-32gb-mlx-dense"
   cat >> maple_server.py <<'MAPLEFILE'
               self.send_response(status);self.send_header('Content-Type','application/json');self.send_header('Content-Length',str(len(data)));self.send_header('Connection','close');self.end_headers();self.wfile.write(data);self.close_connection=True
           def auth(self):
               values=self.headers.get_all('Authorization',[])
               return len(values)==1 and hmac.compare_digest(values[0].encode(),('Bearer '+key).encode())
           def do_GET(self):
               if not self.auth():return self.reply(401,{'error':'unauthorized'})
               if self.path!='/health':return self.reply(404,{'error':'not_found'})
               return self.reply(200,{'ready':True,'backend':mode,'context_limit':CONTEXT,'output_cap':CAP,'head':'dense'})
           def do_POST(self):
               if not self.auth():return self.reply(401,{'error':'unauthorized'})
               if self.path!='/v1/fixed-reasoning-code':return self.reply(404,{'error':'not_found'})
               lengths=self.headers.get_all('Content-Length',[])
               if self.headers.get('Transfer-Encoding')is not None or len(lengths)!=1:return self.reply(400,{'error':'framing'})
               value=lengths[0]
               if not value or len(value)>3 or any(x not in '0123456789'for x in value)or not 1<=int(value)<=128:return self.reply(413,{'error':'body_limit'})
               if self.headers.get_content_type()!='application/json':return self.reply(415,{'error':'json_required'})
               try:
                   body=self.rfile.read(int(value))
                   if len(body)!=int(value):raise ValueError('incomplete')
                   def unique(pairs):
                       d={}
                       for k,v in pairs:
                           if k in d:raise ValueError('duplicate')
                           d[k]=v
                       return d
                   d=json.loads(body.decode(),object_pairs_hook=unique,parse_constant=lambda _:(_ for _ in ()).throw(ValueError('nonfinite')))
                   if type(d)is not dict or d!={'job':'reasoning-code'}:raise ValueError('fixed_job')
               except (ValueError,UnicodeError,TimeoutError,OSError,RecursionError):return self.reply(400,{'error':'invalid_request'})
               try:return self.reply(200,checked_result(mode,engine()))
               except Exception:return self.reply(500,{'error':'native_generation_or_boundary_failure'})
       return HTTPServer(('127.0.0.1',port),Handler)
   if __name__=='__main__':
       if sys.argv[1:]not in (['cuda'],['mlx']):raise SystemExit('Use cuda or mlx')
       mode=sys.argv[1];key=(ROOT/'.api-key').read_text().strip()
       if len(key)!=64 or any(x not in '0123456789abcdef'for x in key):raise SystemExit('private_key_format')
       engine=make_engine(mode)
       with create_server(engine,key,mode)as server:server.serve_forever()
   MAPLEFILE

   Expected result: The exact source section is written. Continue through every section before verification.

5. 5. Bind helper integrity and stable private access

   Generate a key only when absent. Every startup verifies all assets again before remote code/model loading. Keep the key private; it is never printed.

   CommandCopy command

   set -eu
   cd "$HOME/maple-preview-local-m4-32gb-mlx-dense"
   umask 077
   if [ ! -e .api-key ]; then openssl rand -hex 32 > .api-key; fi
   chmod 600 .api-key
   .venv/bin/python - <<'PYCHECK'
   import hashlib,pathlib,ast
   p=pathlib.Path("maple_server.py");assert hashlib.sha256(p.read_bytes()).hexdigest()=="da4d1448e28fd704b8bad98249a657053c055caaec29415951da23966564f0dd"
   ast.parse(p.read_text())
   k=pathlib.Path(".api-key").read_text().strip();assert len(k)==64 and all(c in "0123456789abcdef" for c in k)
   PYCHECK

   Expected result: The complete helper checksum, syntax and stable key format pass without model execution.

Start the model

1. 1. Start the serial native loopback service

   Run in the foreground. Readiness is returned only after exact native assets and full loading checks pass. Native compute has no hard cancellation deadline; close the process explicitly after a failed or timed-out run.

   CommandCopy command

   set -eu
   cd "$HOME/maple-preview-local-m4-32gb-mlx-dense"
   export HF_HUB_OFFLINE=1 TRANSFORMERS_OFFLINE=1 TOKENIZERS_PARALLELISM=false
   .venv/bin/python reader_assets.py mlx verify
   exec .venv/bin/python maple_server.py mlx

   Expected result: After actual reader startup, authenticated health reports the selected backend, dense head and bounded workload. Stop on loading, kernel, platform, memory or checksum failure.

Verify the setup

1. 1. Verify private access and complete native two-turn output

   Use a second terminal on the same machine and the exact virtual-environment Python. The request contains only a fixed job name. The client timeout does not cancel native computation. Record actual output and inspect arithmetic and code; no semantic gold result is forced.

   CommandCopy command

   set -eu
   cd "$HOME/maple-preview-local-m4-32gb-mlx-dense"
   .venv/bin/python - <<'PYVERIFY'
   import urllib.request,urllib.error,pathlib,json,importlib.util
   k=pathlib.Path('.api-key').read_text().strip();base='http://127.0.0.1:9008'
   try:urllib.request.urlopen(base+'/health',timeout=5);raise AssertionError('unauthenticated health accepted')
   except urllib.error.HTTPError as e:assert e.code==401
   def call(path,data=None,timeout=5):
    r=urllib.request.Request(base+path,data=data,headers={'Authorization':'Bearer '+k,'Content-Type':'application/json'})
    with urllib.request.urlopen(r,timeout=timeout)as x:return json.load(x)
   h=call('/health');assert h=={'ready':True,'backend':'mlx','context_limit':8192,'output_cap':4096,'head':'dense'}
   r=call('/v1/fixed-reasoning-code',b'{"job":"reasoning-code"}',timeout=3600)
   s=importlib.util.spec_from_file_location('reader_boundary','maple_server.py');m=importlib.util.module_from_spec(s);s.loader.exec_module(m)
   m.checked_result('mlx',r);pathlib.Path('actual-reader-result.json').write_text(json.dumps(r,indent=2)+'\n')
   for i,t in enumerate(r['turns'],1):print('Turn',i,'input/output tokens',t['input_tokens'],t['output_tokens'],'\nReasoning:',t['reasoning'],'\nFinal:',t['answer'])
   print('Inspect arithmetic 8 then 9 and function update manually; no returned code is executed and no model correctness is assumed.')
   PYVERIFY

   Expected result: Unauthenticated access returns 401. Authenticated health and two complete raw-EOS, closed-thinking turns pass exact token-budget checks. Human inspection determines actual task quality; code is not executed.

Performance evidence

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

Limits and troubleshooting

• Only exact source and metadata plus stdlib fake/AST checks were executed. Native environment, weights, kernels, model quality, memory peak and speed remain unmeasured.
• The five-second socket timeout bounds request framing only. Serial generation can run much longer and requires explicit process shutdown; this is not an inference SLA.
• The fixed reasoning/code inputs do not establish unrestricted tool, production, public API or whole-family coverage. Required guide totals remain null.

Asset, native import, GPU, extension or loading failure

Stop and retain diagnostics without secret values. Do not bypass checks, omit experts, change source math, substitute another checkpoint or enable an approximate head. A different platform or backend requires its own qualification.

Length exhaustion, missing EOS or malformed thinking/control output

Treat the result as incomplete. Preserve actual output and token counts; do not strip native control fragments, fabricate a final answer or execute returned code.

Out of memory or client timeout

Stop the foreground process and verify endpoint closure. Native compute is not cancelled by the socket or client timeout. Record actual failure; no fit or latency was measured.

Stop the model

1. 1. Stop and verify endpoint closure

   Press Ctrl-C in the foreground service terminal. Then verify the private endpoint is closed.

   CommandCopy command

   set -eu
   cd "$HOME/maple-preview-local-m4-32gb-mlx-dense"
   .venv/bin/python - <<'PYSTOP'
   import socket
   with socket.socket()as s:
    s.settimeout(3);assert s.connect_ex(('127.0.0.1',9008))!=0
   print('Private service endpoint is closed.')
   PYSTOP

   Expected result: The native Python process has stopped and the loopback endpoint is closed.

Uninstall

1. 1. Remove only this named reader workspace

   After stopping and saving any needed result/version records, remove this exact workspace and its downloaded native assets/environment. Do not remove shared system drivers or unrelated files.

   CommandCopy command

   set -eu
   python3 - <<'PYCLEAN'
   from pathlib import Path
   import shutil
   p=Path.home()/"maple-preview-local-m4-32gb-mlx-dense"
   assert p.name=="maple-preview-local-m4-32gb-mlx-dense" and p.parent==Path.home() and not p.is_symlink()
   if p.exists():shutil.rmtree(p)
   PYCLEAN

   Expected result: Only the named private workspace is removed; shared machine software remains separate.

Exact model and requirements

Artifact:

DeepGrove AI / Maple Preview complete native BF16 CUDA checkpoint · checkpoint deepgrove/maple-preview

Publisher revision:

ac1ddd79d2b5cb4406f5d2bebdf95406ce505a07

Artifact digest:

No composite whole-package digest is published. Every required file has exact size and LFS SHA-256 or native Git blob SHA-1 in the embedded mandatory manifest.

Download size:

37.65 GiB

Memory requirement (unknown):

Unknown — No complete native runtime, cache, activation, loading-buffer or OS peak was measured.

System RAM / disk:

Unknown / Unknown

Placement:

single device. One selected accelerator; all layers and experts remain present. Complete model and caches use one H100; no CPU or disk weight offload. Host loading buffers remain additional.

KV cache:

Native BF16 attention KV on the selected CUDA model path; exact realized peak remains unmeasured

Workload:

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

Tokenizer:

Exact native Maple tokenizer and 3,292-byte publisher Jinja, with no added BOS · ac1ddd79d2b5cb4406f5d2bebdf95406ce505a07

Sampling / reasoning:

Dense full-vocabulary greedy argmax is an author-selected fixed-task setting. Native direct forward and causal/sliding FA2 mathematics are unchanged. The publisher template opens thinking. Preserve complete reasoning_content and final answer in the second turn. Require exactly closed thinking and true raw native EOS; reject partial/reopened/tool/control output. 8,192 total context and 4,096 output cap are walkthrough bounds, not the native full-context recommendation.

Runtime prerequisites:

Ubuntu 24.04 x86_64; NVIDIA driver >= 550.54.14 for CUDA 12.4; working H100 visibility, Exact FA2 wheel ABI selected from actual Torch build; FA3 and Liger absent, Python 3.12.

Official artifact · License

Exact publisher MIT license is archived. Retain its copyright and permission notice in copies or substantial portions; software is provided without warranty or liability. The separate publisher MLX fork is also MIT. No terms are accepted or legal clearance granted by this guide.

• LICENSE · 0 GiB · revision ac1ddd79d2b5cb4406f5d2bebdf95406ce505a07
• README.md · 0 GiB · revision ac1ddd79d2b5cb4406f5d2bebdf95406ce505a07
• added_tokens.json · 0 GiB · revision ac1ddd79d2b5cb4406f5d2bebdf95406ce505a07
• chat_template.jinja · 0 GiB · revision ac1ddd79d2b5cb4406f5d2bebdf95406ce505a07
• config.json · 0 GiB · revision ac1ddd79d2b5cb4406f5d2bebdf95406ce505a07
• configuration_maple.py · 0 GiB · revision ac1ddd79d2b5cb4406f5d2bebdf95406ce505a07
• fa3.py · 0 GiB · revision ac1ddd79d2b5cb4406f5d2bebdf95406ce505a07
• merges.txt · 0 GiB · revision ac1ddd79d2b5cb4406f5d2bebdf95406ce505a07
• model.safetensors.index.json · 0 GiB · revision ac1ddd79d2b5cb4406f5d2bebdf95406ce505a07
• modeling_maple.py · 0 GiB · revision ac1ddd79d2b5cb4406f5d2bebdf95406ce505a07
• special_tokens_map.json · 0 GiB · revision ac1ddd79d2b5cb4406f5d2bebdf95406ce505a07
• tokenizer.json · 0.01 GiB · revision ac1ddd79d2b5cb4406f5d2bebdf95406ce505a07
• tokenizer_config.json · 0 GiB · revision ac1ddd79d2b5cb4406f5d2bebdf95406ce505a07
• vocab.json · 0 GiB · revision ac1ddd79d2b5cb4406f5d2bebdf95406ce505a07

• 40430372376 serialized weight bytes do not include complete runtime memory.
• All experts, caches, native extension workspaces, activations, loading buffers and OS overhead remain additional and unmeasured.

Install the runtime

1. 1. Prepare the existing reader machine

   Use the exact platform in prerequisites. Record actual capacity before downloading. These are planning checks, not measured model fit.

   CommandCopy command

   set -eu
   mkdir -p "$HOME/maple-preview-aws-p5-4xlarge-bf16-native"
   cd "$HOME/maple-preview-aws-p5-4xlarge-bf16-native"
   sudo apt-get update
   sudo apt-get install -y python3.12-venv python3.12 curl ca-certificates openssl git
   nvidia-smi --query-gpu=name,memory.total,driver_version --format=csv,noheader
   python3.12 -m venv .venv
   python3.12 - <<'PYPREFLIGHT'
   import shutil
   assert shutil.disk_usage('.').free >= 171798691840
   print('Reader prescribed free space passed; fit remains unmeasured.')
   PYPREFLIGHT

   Expected result: An existing compatible machine, Python 3.12 environment and prescribed free space are recorded. Stop on platform, GPU, driver or disk failure.

Download the model

1. 1. Write the mandatory streaming asset checker (1/1)

   Write this exact source section in order. Writing source does not execute native code.

   CommandCopy command

   set -eu
   cd "$HOME/maple-preview-aws-p5-4xlarge-bf16-native"
   cat > reader_assets.py <<'MAPLEFILE'
   """Reader-only exact full Maple asset verification; no native imports."""
   import hashlib,json,pathlib,sys,urllib.request
   ROOT=pathlib.Path(__file__).resolve().parent
   PINS={'cuda':('deepgrove/maple-preview','ac1ddd79d2b5cb4406f5d2bebdf95406ce505a07','3ab6ce508a608d0af564c6683497eb7d166178d8c992e33dd41404c9e2be92eb',23),'mlx':('deepgrove/maple-preview-2bit-mlx','e8be321fb325c40beb7b621bd98e8e57a324ec42','d236df54f11ada042e099c2ae15942b246c611420495ee39ac3a470a38b5a182',15)}
   def manifest(mode):
       repo,pin,digest,count=PINS[mode];data=(ROOT/'assets.json').read_bytes()
       if hashlib.sha256(data).hexdigest()!=digest:raise ValueError('manifest_hash')
       m=json.loads(data)
       if m['mode']!=mode or m['repo']!=repo or m['revision']!=pin or len(m['files'])!=count:raise ValueError('manifest_identity')
       return m
   
   def check(path,row):
       h=hashlib.sha256()if row['algorithm']=='sha256'else hashlib.sha1();n=0
       if row['algorithm']=='git-sha1':h.update(b'blob '+str(row['bytes']).encode()+b'\0')
       with path.open('rb')as f:
           for b in iter(lambda:f.read(1024*1024),b''):n+=len(b);h.update(b)
       if n!=row['bytes']or h.hexdigest()!=row['digest']:raise ValueError('asset_integrity_'+row['path'])
   
   def verify(mode,download=False):
       m=manifest(mode)
       for category,dirname in [('files','model'),('runtime_files','runtime')]:
           root=ROOT/dirname;root.mkdir(exist_ok=True)
           for row in m[category]:
               if pathlib.PurePosixPath(row['path']).name!=row['path']:raise ValueError('asset_path')
               path=root/row['path']
               if download and not path.exists():
                   temp=path.with_suffix(path.suffix+'.partial');n=0
                   with urllib.request.urlopen(row['url'],timeout=120)as src,temp.open('wb')as dst:
                       for chunk in iter(lambda:src.read(1024*1024),b''):
                           n+=len(chunk)
                           if n>row['bytes']:raise ValueError('download_size')
                           dst.write(chunk)
                   check(temp,row);temp.replace(path)
               check(path,row)
       index=json.loads((ROOT/'model/model.safetensors.index.json').read_text())['weight_map'];parts={x['path']for x in m['files']if x['path'].endswith('.safetensors')}
       # Publisher MLX index covers the three main parts; FlashHead is a mandatory
       # separately shipped fourth asset, loaded by publisher model*.safetensors glob.
       indexed=parts-({'model-flashhead.safetensors'}if mode=='mlx'else set())
       if set(index.values())!=indexed:raise ValueError('complete_weight_index')
       print('All exact native or publisher-packed files verified; no fit or numerical equivalence proof.')
   if __name__=='__main__':
       if len(sys.argv)!=3 or sys.argv[1]not in PINS or sys.argv[2]not in ('download','verify'):raise SystemExit('Use cuda|mlx download|verify')
       verify(sys.argv[1],sys.argv[2]=='download')
   MAPLEFILE

   Expected result: The exact source section is written. Continue through every section before verification.

2. 2. Write the exact complete publisher asset manifest (1/2)

   Write this exact source section in order. Writing source does not execute native code.

   CommandCopy command

   set -eu
   cd "$HOME/maple-preview-aws-p5-4xlarge-bf16-native"
   cat > assets.json <<'MAPLEFILE'
   MAPLEFILE

   Expected result: The exact source section is written. Continue through every section before verification.

3. 3. Write the exact complete publisher asset manifest (2/2)

   Write this exact source section in order. Writing source does not execute native code.

   CommandCopy command

   set -eu
   cd "$HOME/maple-preview-aws-p5-4xlarge-bf16-native"
   cat >> assets.json <<'MAPLEFILE'
   {"mode":"cuda","repo":"deepgrove/maple-preview","revision":"ac1ddd79d2b5cb4406f5d2bebdf95406ce505a07","files":[{"path":"LICENSE","bytes":1066,"algorithm":"git-sha1","digest":"26a36c1b035374c2e22090ddc52c1a00a727ecb6","url":"https://huggingface.co/deepgrove/maple-preview/resolve/ac1ddd79d2b5cb4406f5d2bebdf95406ce505a07/LICENSE"},{"path":"README.md","bytes":1941,"algorithm":"git-sha1","digest":"70a6cdf094695e092f3b9e62475ab8a17221a68e","url":"https://huggingface.co/deepgrove/maple-preview/resolve/ac1ddd79d2b5cb4406f5d2bebdf95406ce505a07/README.md"},{"path":"added_tokens.json","bytes":707,"algorithm":"git-sha1","digest":"b54f9135e44c1e81047e8d05cb027af8bc039eed","url":"https://huggingface.co/deepgrove/maple-preview/resolve/ac1ddd79d2b5cb4406f5d2bebdf95406ce505a07/added_tokens.json"},{"path":"chat_template.jinja","bytes":3292,"algorithm":"git-sha1","digest":"b5509b8aef363d7262fb86370db1eb41f554959b","url":"https://huggingface.co/deepgrove/maple-preview/resolve/ac1ddd79d2b5cb4406f5d2bebdf95406ce505a07/chat_template.jinja"},{"path":"config.json","bytes":1893,"algorithm":"git-sha1","digest":"679618c2e5d73619b9809dfbe77e5e049f1db99c","url":"https://huggingface.co/deepgrove/maple-preview/resolve/ac1ddd79d2b5cb4406f5d2bebdf95406ce505a07/config.json"},{"path":"configuration_maple.py","bytes":2149,"algorithm":"git-sha1","digest":"d78f1d1b06f397f02192a8c157d71c1e0fe9d91c","url":"https://huggingface.co/deepgrove/maple-preview/resolve/ac1ddd79d2b5cb4406f5d2bebdf95406ce505a07/configuration_maple.py"},{"path":"fa3.py","bytes":9681,"algorithm":"git-sha1","digest":"cf5d5d07a7076d6cdb1f541e929b67e084c97cb0","url":"https://huggingface.co/deepgrove/maple-preview/resolve/ac1ddd79d2b5cb4406f5d2bebdf95406ce505a07/fa3.py"},{"path":"merges.txt","bytes":1671853,"algorithm":"git-sha1","digest":"31349551d90c7606f325fe0f11bbb8bd5fa0d7c7","url":"https://huggingface.co/deepgrove/maple-preview/resolve/ac1ddd79d2b5cb4406f5d2bebdf95406ce505a07/merges.txt"},{"path":"model-00001-of-00009.safetensors","bytes":4998990664,"algorithm":"sha256","digest":"a0e75504d43e41c5aa1796abac3266e8626706742f2c6da8a67a48d31d9a4f77","url":"https://huggingface.co/deepgrove/maple-preview/resolve/ac1ddd79d2b5cb4406f5d2bebdf95406ce505a07/model-00001-of-00009.safetensors"},{"path":"model-00002-of-00009.safetensors","bytes":4999904624,"algorithm":"sha256","digest":"166113ab0c8f485d5d8a6060c4f1714ed4a7103ba703fd5c51316fb70c5abdd6","url":"https://huggingface.co/deepgrove/maple-preview/resolve/ac1ddd79d2b5cb4406f5d2bebdf95406ce505a07/model-00002-of-00009.safetensors"},{"path":"model-00003-of-00009.safetensors","bytes":4999904624,"algorithm":"sha256","digest":"82c08761b18236f95e06b04ebb24ef55e2a54d0abda95afe6ee8ad963ce712dd","url":"https://huggingface.co/deepgrove/maple-preview/resolve/ac1ddd79d2b5cb4406f5d2bebdf95406ce505a07/model-00003-of-00009.safetensors"},{"path":"model-00004-of-00009.safetensors","bytes":4999905768,"algorithm":"sha256","digest":"dda74a69c9ac24dfaf135b0d459b47307e134dc5dc2cd7c4171f249a6c267fa0","url":"https://huggingface.co/deepgrove/maple-preview/resolve/ac1ddd79d2b5cb4406f5d2bebdf95406ce505a07/model-00004-of-00009.safetensors"},{"path":"model-00005-of-00009.safetensors","bytes":4999907000,"algorithm":"sha256","digest":"73b5a1edfcc317374305ea29c79b0ddb184f8f069163f9083bffe0b0181a00ec","url":"https://huggingface.co/deepgrove/maple-preview/resolve/ac1ddd79d2b5cb4406f5d2bebdf95406ce505a07/model-00005-of-00009.safetensors"},{"path":"model-00006-of-00009.safetensors","bytes":4999907000,"algorithm":"sha256","digest":"f8533087e0a88695d8bb43fa82d2be16e396e8fbc1e32357d8cfa99af6860446","url":"https://huggingface.co/deepgrove/maple-preview/resolve/ac1ddd79d2b5cb4406f5d2bebdf95406ce505a07/model-00006-of-00009.safetensors"},{"path":"model-00007-of-00009.safetensors","bytes":4999906992,"algorithm":"sha256","digest":"efcb94993cb5d9062c879c4d132489f58687b5e21ff20d2913b011d6504127dd","url":"https://huggingface.co/deepgrove/maple-preview/resolve/ac1ddd79d2b5cb4406f5d2bebdf95406ce505a07/model-00007-of-00009.safetensors"},{"path":"model-00008-of-00009.safetensors","bytes":4999907000,"algorithm":"sha256","digest":"79be22027606cf7efde7d6f409a6cf369380132d1baf625bcc90f4280a7baab0","url":"https://huggingface.co/deepgrove/maple-preview/resolve/ac1ddd79d2b5cb4406f5d2bebdf95406ce505a07/model-00008-of-00009.safetensors"},{"path":"model-00009-of-00009.safetensors","bytes":432038704,"algorithm":"sha256","digest":"4761e38c77d93f4807181eb972557f44649336c0d261064578a9d3c1587cbebb","url":"https://huggingface.co/deepgrove/maple-preview/resolve/ac1ddd79d2b5cb4406f5d2bebdf95406ce505a07/model-00009-of-00009.safetensors"},{"path":"model.safetensors.index.json","bytes":1685937,"algorithm":"git-sha1","digest":"371d69d9b4c178f2bd1efc223f394d5f6e2e81c9","url":"https://huggingface.co/deepgrove/maple-preview/resolve/ac1ddd79d2b5cb4406f5d2bebdf95406ce505a07/model.safetensors.index.json"},{"path":"modeling_maple.py","bytes":28642,"algorithm":"git-sha1","digest":"d33d1aedfa2bdbd4defea9123c3f2b9b56f5ed21","url":"https://huggingface.co/deepgrove/maple-preview/resolve/ac1ddd79d2b5cb4406f5d2bebdf95406ce505a07/modeling_maple.py"},{"path":"special_tokens_map.json","bytes":613,"algorithm":"git-sha1","digest":"ac23c0aaa2434523c494330aeb79c58395378103","url":"https://huggingface.co/deepgrove/maple-preview/resolve/ac1ddd79d2b5cb4406f5d2bebdf95406ce505a07/special_tokens_map.json"},{"path":"tokenizer.json","bytes":11422654,"algorithm":"sha256","digest":"aeb13307a71acd8fe81861d94ad54ab689df773318809eed3cbe794b4492dae4","url":"https://huggingface.co/deepgrove/maple-preview/resolve/ac1ddd79d2b5cb4406f5d2bebdf95406ce505a07/tokenizer.json"},{"path":"tokenizer_config.json","bytes":5432,"algorithm":"git-sha1","digest":"f4b91854389fd13e964c436e576a0011e588104b","url":"https://huggingface.co/deepgrove/maple-preview/resolve/ac1ddd79d2b5cb4406f5d2bebdf95406ce505a07/tokenizer_config.json"},{"path":"vocab.json","bytes":2776833,"algorithm":"git-sha1","digest":"4783fe10ac3adce15ac8f358ef5462739852c569","url":"https://huggingface.co/deepgrove/maple-preview/resolve/ac1ddd79d2b5cb4406f5d2bebdf95406ce505a07/vocab.json"}],"runtime_files":[]}
   MAPLEFILE

   Expected result: The exact source section is written. Continue through every section before verification.

4. 4. Verify the manifest and download every required asset

   Only the reader runs this download. All model parts, tokenizer auxiliaries, remote code and the Apple kernel wheel are mandatory. Existing mismatched files fail closed; no byte range or alternate checkpoint is used.

   CommandCopy command

   set -eu
   cd "$HOME/maple-preview-aws-p5-4xlarge-bf16-native"
   .venv/bin/python - <<'PYHASH'
   import hashlib,pathlib
   for n,h in [('reader_assets.py', 'e12f5b1a39737ba1649b235702ebe76404406d7422e1593c3a20b20a34fe56fe'), ('assets.json', '3ab6ce508a608d0af564c6683497eb7d166178d8c992e33dd41404c9e2be92eb')]:
    assert hashlib.sha256(pathlib.Path(n).read_bytes()).hexdigest()==h
   PYHASH
   .venv/bin/python reader_assets.py cuda download

   Expected result: All required per-file sizes and streaming hashes pass before any native loading. This establishes exact bytes, not model quality or fit.

Configure

1. 1. Install the exact native-facing reader runtime

   Reader installation is source-qualified and unexecuted here. Use a fresh environment; no FA3, Liger, alternate MLX fork or floating kernel build. Stop on wheel ABI, dependency or native import failure. Save actual transitive versions; this is not an executed lockfile.

   CommandCopy command

   set -eu
   cd "$HOME/maple-preview-aws-p5-4xlarge-bf16-native"
   .venv/bin/python -m pip install "https://download-r2.pytorch.org/whl/cu124/torch-2.6.0%2Bcu124-cp312-cp312-linux_x86_64.whl#sha256=a393b506844035c0dac2f30ea8478c343b8e95a429f06f3b3cadfc7f53adb597" transformers==4.57.1 huggingface-hub==0.36.0 tokenizers==0.22.2 safetensors==0.5.3 numpy==2.2.3 einops==0.8.1 PyYAML==6.0.2 Jinja2==3.1.6
   .venv/bin/python - <<'PYFA'
   import torch,urllib.request,hashlib,pathlib,subprocess,sys
   assert torch.__version__.split("+")[0]=="2.6.0" and torch.version.cuda=="12.4"
   abi=bool(torch._C._GLIBCXX_USE_CXX11_ABI)
   label="TRUE" if abi else "FALSE"
   h={False:"51100460844b0cb4bd5adc38af60c56adff8904508861aeb5ef7a1c2e41f529a",True:"ae3b8b402ca649e0974f20a9b8641cc5a72233722674cf0f7b524fbdf736fdb7"}[abi]
   name="flash_attn-2.8.3+cu12torch2.6cxx11abi"+label+"-cp312-cp312-linux_x86_64.whl"
   u="https://github.com/Dao-AILab/flash-attention/releases/download/v2.8.3/"+name.replace("+","%2B")
   p=pathlib.Path(name);digest=hashlib.sha256()
   with urllib.request.urlopen(u,timeout=120)as r,p.open("wb")as f:
    while b:=r.read(1024*1024):digest.update(b);f.write(b)
   assert digest.hexdigest()==h
   subprocess.run([sys.executable,"-m","pip","install","--no-deps",str(p)],check=True)
   PYFA
   .venv/bin/python -m pip check
   .venv/bin/python -m pip freeze > realized-environment.txt

   Expected result: Exact native-facing versions install and pip check passes only after reader execution. Record realized versions; no compatibility or fit was measured by the author.

2. 2. Write the complete private native reasoning service (1/3)

   Write this exact source section in order. Writing source does not execute native code.

   CommandCopy command

   set -eu
   cd "$HOME/maple-preview-aws-p5-4xlarge-bf16-native"
   cat > maple_server.py <<'MAPLEFILE'
   """Reader-authored fixed Maple reasoning/code harness. Native imports occur only in reader startup."""
   import hashlib,hmac,json,pathlib,sys
   from http.server import BaseHTTPRequestHandler,HTTPServer
   ROOT=pathlib.Path(__file__).resolve().parent
   CONTEXT=8192;CAP=4096;EOS=151645;OPEN=151667;CLOSE=151668
   PROMPTS=['A library opens at 09:00 and closes at 17:00 on the same day. Work out its opening duration, then give a small Python function opening_hours() returning that integer. Do not execute code or call tools.','Now closing changes to 18:00 with the same opening time. Recompute and update the function. Do not execute code or call tools.']
   def native_forward(model,input_ids,cache):
       # Exact native no-mask, no-explicit-position path. Never model.generate().
       return model(input_ids=input_ids,attention_mask=None,position_ids=None,
           past_key_values=cache,use_cache=True,cache_position=None,return_dict=True,logits_to_keep=1)
   
   def parse_output(ids,text):
       if type(ids)is not list or not 1<=len(ids)<=CAP or any(type(x)is not int or x<0 for x in ids):raise ValueError('raw_output_ids')
       if ids[-1]!=EOS or ids.count(EOS)!=1 or ids.count(CLOSE)!=1 or OPEN in ids:raise ValueError('native_end_or_think')
       end=ids.index(CLOSE)
       if any(x>=151643 and x not in (OPEN,CLOSE,EOS)for x in ids):raise ValueError('unexpected_native_control_id')
       if type(text)is not str or text.count('</think>')!=1:raise ValueError('decoded_thinking_boundary')
       reasoning,answer=text.split('</think>')
       if not reasoning.strip()or not answer.strip()or any(x in reasoning+answer for x in ('<think','</think','<tool','</tool','<|','|>')):raise ValueError('nonfinal_control_or_empty')
       return reasoning,answer
   
   def checked_result(mode,d):
       if type(d)is not dict or set(d)!={'backend','turns','context_limit','output_cap','head'}or d['backend']!=mode or d['context_limit']!=CONTEXT or d['output_cap']!=CAP or d['head']!='dense':raise ValueError('result_identity')
       if type(d['turns'])is not list or len(d['turns'])!=2:raise ValueError('two_turns_required')
       for r in d['turns']:
           if type(r)is not dict or set(r)!={'reasoning','answer','input_tokens','output_tokens','stop_id'}:raise ValueError('turn_schema')
           if any(type(r[k])is not int for k in ('input_tokens','output_tokens','stop_id'))or not 1<=r['input_tokens']<=CONTEXT-CAP or not 1<=r['output_tokens']<=CAP or r['input_tokens']+r['output_tokens']>CONTEXT or r['stop_id']!=EOS:raise ValueError('turn_budget_stop')
           if any(type(r[k])is not str or not r[k].strip() or len(r[k])>65536 for k in ('reasoning','answer')):raise ValueError('turn_text')
           if any(c in r['reasoning']+r['answer']for c in ('<think','</think','<tool','</tool','<|','|>')):raise ValueError('turn_control')
       return d
   
   def publisher_native_extension(model):
       # Publisher loader executes custom_model without registering sys.modules.
       # The verified source-defined class initializer retains that module globals.
       init=model.__class__.__init__
       context=getattr(init,'__globals__',None)
       if type(context)is not dict or context.get('__name__')!=model.__class__.__module__:
           raise RuntimeError('publisher_loaded_context_required')
       native=context.get('_maple_native')
       if native is None:raise RuntimeError('pinned_publisher_kernel_required')
       return native
   
   def make_engine(mode):
       from reader_assets import verify
       verify(mode)
       path=ROOT/'model';cfg=json.loads((path/'config.json').read_text())
       if cfg['num_hidden_layers']!=24 or cfg['num_experts']!=256 or cfg['num_experts_per_tok']!=8 or cfg['sliding_window']!=512:raise RuntimeError('wrong_maple_architecture')
       if mode=='cuda':
           import torch
           from transformers import AutoTokenizer,AutoModelForCausalLM
           if torch.__version__.split('+')[0]!='2.6.0'or torch.version.cuda!='12.4'or torch.cuda.device_count()!=1 or 'H100'not in torch.cuda.get_device_name(0):raise RuntimeError('exact_cuda_runtime_required')
           # Select only the explicit FA2 fallback; FA3 is deliberately not installed.
           import importlib.util
           if importlib.util.find_spec('flash_attn_interface')is not None:raise RuntimeError('unexpected_fa3_interface')
           import flash_attn
           if flash_attn.__version__!='2.8.3':raise RuntimeError('flash_attention_version')
           import inspect
           if 'window_size'not in inspect.signature(flash_attn.flash_attn_func).parameters or importlib.util.find_spec('liger_kernel')is not None:raise RuntimeError('exact_unfused_sliding_attention_backend')
           if cfg['max_position_embeddings']!=131072 or cfg['nope_on_global_attention']is not True:raise RuntimeError('native_cuda_config')
           tok=AutoTokenizer.from_pretrained(path,local_files_only=True,trust_remote_code=False)
   MAPLEFILE

   Expected result: The exact source section is written. Continue through every section before verification.

3. 3. Write the complete private native reasoning service (2/3)

   Write this exact source section in order. Writing source does not execute native code.

   CommandCopy command

   set -eu
   cd "$HOME/maple-preview-aws-p5-4xlarge-bf16-native"
   cat >> maple_server.py <<'MAPLEFILE'
           model,info=AutoModelForCausalLM.from_pretrained(path,local_files_only=True,trust_remote_code=True,torch_dtype=torch.bfloat16,device_map=None,use_safetensors=True,output_loading_info=True,ignore_mismatched_sizes=False)
           keys={'missing_keys','unexpected_keys','mismatched_keys','error_msgs'}
           if type(info)is not dict or set(info)!=keys or any(info[k]for k in keys):raise RuntimeError('native_load_diagnostics')
           model=model.to('cuda:0').eval()
           if len(model.model.layers)!=24 or any(p.dtype!=torch.bfloat16 or p.device.type!='cuda' or not torch.isfinite(p).all()for p in model.parameters()):raise RuntimeError('full_native_gpu_parameters')
           def generate(ids):
               tokens=[];cache=None;current=torch.tensor([ids],dtype=torch.long,device='cuda:0')
               with torch.inference_mode():
                   for _ in range(CAP):
                       out=native_forward(model,current,cache);cache=out.past_key_values;scores=out.logits[0,-1]
                       if not torch.isfinite(scores).all():raise RuntimeError('nonfinite_native_logits')
                       token=int(torch.argmax(scores));tokens.append(token)
                       if token==EOS:break
                       current=torch.tensor([[token]],dtype=torch.long,device='cuda:0')
               return tokens,tok.decode(tokens[:-1]if tokens and tokens[-1]==EOS else tokens,skip_special_tokens=False)
       elif mode=='mlx':
           import platform,subprocess
           if platform.machine()!='arm64'or tuple(map(int,platform.mac_ver()[0].split('.')[:2]))<(26,2):raise RuntimeError('native_apple_platform')
           import mlx.core as mx
           from mlx_lm import load
           from mlx_lm.generate import stream_generate
           if mx.__version__!='0.32.0'or mx.default_device()!=mx.gpu:raise RuntimeError('exact_mlx_metal_runtime')
           if cfg['max_position_embeddings']!=128000 or cfg.get('nope_on_global_attention')is not False:raise RuntimeError('publisher_mlx_config')
           model,tok=load(str(path),model_config={'use_flash_head':False},trust_remote_code=True)
           if model.lm_head_flash is not None or len(model.model.layers)!=24:raise RuntimeError('dense_complete_maple_required')
           import importlib.metadata
           native=publisher_native_extension(model)
           if importlib.metadata.version('mlx-lm-maple-kernels')!='0.1.0'or native is None:raise RuntimeError('pinned_publisher_kernel_required')
           def sampler(logprobs):
               if bool(mx.any(mx.isnan(logprobs)))or bool(mx.any(mx.isinf(logprobs)&(logprobs>0)))or not bool(mx.any(mx.isfinite(logprobs))):raise RuntimeError('nonfinite_dense_logits')
               return mx.argmax(logprobs,axis=-1)
           def generate(ids):
               tokens=[];last=None
               for item in stream_generate(model,tok,ids,max_tokens=CAP,sampler=sampler):
                   tokens.append(item.token);last=item
               if last is None or last.finish_reason!='stop' or last.generation_tokens!=len(tokens):raise RuntimeError('native_mlx_completion')
               return tokens,tok.decode(tokens[:-1]if tokens and tokens[-1]==EOS else tokens,skip_special_tokens=False)
       else:raise ValueError('backend')
       if tok.eos_token_id!=EOS or tok.convert_tokens_to_ids('</think>')!=CLOSE:raise RuntimeError('native_token_identity')
       def engine():
           messages=[];turns=[]
           for prompt in PROMPTS:
               messages.append({'role':'user','content':prompt})
               rendered=tok.apply_chat_template(messages,tokenize=False,add_generation_prompt=True)
               ids=tok.encode(rendered,add_special_tokens=False)
               if not 1<=len(ids)<=CONTEXT-CAP:raise ValueError('full_native_history_budget')
               raw,text=generate(ids);reasoning,answer=parse_output(raw,text)
               turns.append({'reasoning':reasoning,'answer':answer,'input_tokens':len(ids),'output_tokens':len(raw),'stop_id':raw[-1]})
               messages.append({'role':'assistant','content':answer,'reasoning_content':reasoning})
           return checked_result(mode,{'backend':mode,'turns':turns,'context_limit':CONTEXT,'output_cap':CAP,'head':'dense'})
       return engine
   
   def create_server(engine,key,mode,port=9008):
       if type(key)is not str or len(key)!=64 or any(x not in '0123456789abcdef'for x in key):raise ValueError('key_format')
       class Handler(BaseHTTPRequestHandler):
           def setup(self):super().setup();self.connection.settimeout(5)
           def log_message(self,*_):pass
           def reply(self,status,d):
               data=json.dumps(d,allow_nan=False).encode()
               if len(data)>131072:status=500;data=b'{"error":"response_limit"}'
   MAPLEFILE

   Expected result: The exact source section is written. Continue through every section before verification.

4. 4. Write the complete private native reasoning service (3/3)

   Write this exact source section in order. Writing source does not execute native code.

   CommandCopy command

   set -eu
   cd "$HOME/maple-preview-aws-p5-4xlarge-bf16-native"
   cat >> maple_server.py <<'MAPLEFILE'
               self.send_response(status);self.send_header('Content-Type','application/json');self.send_header('Content-Length',str(len(data)));self.send_header('Connection','close');self.end_headers();self.wfile.write(data);self.close_connection=True
           def auth(self):
               values=self.headers.get_all('Authorization',[])
               return len(values)==1 and hmac.compare_digest(values[0].encode(),('Bearer '+key).encode())
           def do_GET(self):
               if not self.auth():return self.reply(401,{'error':'unauthorized'})
               if self.path!='/health':return self.reply(404,{'error':'not_found'})
               return self.reply(200,{'ready':True,'backend':mode,'context_limit':CONTEXT,'output_cap':CAP,'head':'dense'})
           def do_POST(self):
               if not self.auth():return self.reply(401,{'error':'unauthorized'})
               if self.path!='/v1/fixed-reasoning-code':return self.reply(404,{'error':'not_found'})
               lengths=self.headers.get_all('Content-Length',[])
               if self.headers.get('Transfer-Encoding')is not None or len(lengths)!=1:return self.reply(400,{'error':'framing'})
               value=lengths[0]
               if not value or len(value)>3 or any(x not in '0123456789'for x in value)or not 1<=int(value)<=128:return self.reply(413,{'error':'body_limit'})
               if self.headers.get_content_type()!='application/json':return self.reply(415,{'error':'json_required'})
               try:
                   body=self.rfile.read(int(value))
                   if len(body)!=int(value):raise ValueError('incomplete')
                   def unique(pairs):
                       d={}
                       for k,v in pairs:
                           if k in d:raise ValueError('duplicate')
                           d[k]=v
                       return d
                   d=json.loads(body.decode(),object_pairs_hook=unique,parse_constant=lambda _:(_ for _ in ()).throw(ValueError('nonfinite')))
                   if type(d)is not dict or d!={'job':'reasoning-code'}:raise ValueError('fixed_job')
               except (ValueError,UnicodeError,TimeoutError,OSError,RecursionError):return self.reply(400,{'error':'invalid_request'})
               try:return self.reply(200,checked_result(mode,engine()))
               except Exception:return self.reply(500,{'error':'native_generation_or_boundary_failure'})
       return HTTPServer(('127.0.0.1',port),Handler)
   if __name__=='__main__':
       if sys.argv[1:]not in (['cuda'],['mlx']):raise SystemExit('Use cuda or mlx')
       mode=sys.argv[1];key=(ROOT/'.api-key').read_text().strip()
       if len(key)!=64 or any(x not in '0123456789abcdef'for x in key):raise SystemExit('private_key_format')
       engine=make_engine(mode)
       with create_server(engine,key,mode)as server:server.serve_forever()
   MAPLEFILE

   Expected result: The exact source section is written. Continue through every section before verification.

5. 5. Bind helper integrity and stable private access

   Generate a key only when absent. Every startup verifies all assets again before remote code/model loading. Keep the key private; it is never printed.

   CommandCopy command

   set -eu
   cd "$HOME/maple-preview-aws-p5-4xlarge-bf16-native"
   umask 077
   if [ ! -e .api-key ]; then openssl rand -hex 32 > .api-key; fi
   chmod 600 .api-key
   .venv/bin/python - <<'PYCHECK'
   import hashlib,pathlib,ast
   p=pathlib.Path("maple_server.py");assert hashlib.sha256(p.read_bytes()).hexdigest()=="da4d1448e28fd704b8bad98249a657053c055caaec29415951da23966564f0dd"
   ast.parse(p.read_text())
   k=pathlib.Path(".api-key").read_text().strip();assert len(k)==64 and all(c in "0123456789abcdef" for c in k)
   PYCHECK

   Expected result: The complete helper checksum, syntax and stable key format pass without model execution.

Start the model

1. 1. Start the serial native loopback service

   Run in the foreground. Readiness is returned only after exact native assets and full loading checks pass. Native compute has no hard cancellation deadline; close the process explicitly after a failed or timed-out run.

   CommandCopy command

   set -eu
   cd "$HOME/maple-preview-aws-p5-4xlarge-bf16-native"
   export HF_HUB_OFFLINE=1 TRANSFORMERS_OFFLINE=1 TOKENIZERS_PARALLELISM=false
   export CUDA_VISIBLE_DEVICES=0
   .venv/bin/python reader_assets.py cuda verify
   exec .venv/bin/python maple_server.py cuda

   Expected result: After actual reader startup, authenticated health reports the selected backend, dense head and bounded workload. Stop on loading, kernel, platform, memory or checksum failure.

Verify the setup

1. 1. Verify private access and complete native two-turn output

   Use a second terminal on the same machine and the exact virtual-environment Python. The request contains only a fixed job name. The client timeout does not cancel native computation. Record actual output and inspect arithmetic and code; no semantic gold result is forced.

   CommandCopy command

   set -eu
   cd "$HOME/maple-preview-aws-p5-4xlarge-bf16-native"
   .venv/bin/python - <<'PYVERIFY'
   import urllib.request,urllib.error,pathlib,json,importlib.util
   k=pathlib.Path('.api-key').read_text().strip();base='http://127.0.0.1:9008'
   try:urllib.request.urlopen(base+'/health',timeout=5);raise AssertionError('unauthenticated health accepted')
   except urllib.error.HTTPError as e:assert e.code==401
   def call(path,data=None,timeout=5):
    r=urllib.request.Request(base+path,data=data,headers={'Authorization':'Bearer '+k,'Content-Type':'application/json'})
    with urllib.request.urlopen(r,timeout=timeout)as x:return json.load(x)
   h=call('/health');assert h=={'ready':True,'backend':'cuda','context_limit':8192,'output_cap':4096,'head':'dense'}
   r=call('/v1/fixed-reasoning-code',b'{"job":"reasoning-code"}',timeout=3600)
   s=importlib.util.spec_from_file_location('reader_boundary','maple_server.py');m=importlib.util.module_from_spec(s);s.loader.exec_module(m)
   m.checked_result('cuda',r);pathlib.Path('actual-reader-result.json').write_text(json.dumps(r,indent=2)+'\n')
   for i,t in enumerate(r['turns'],1):print('Turn',i,'input/output tokens',t['input_tokens'],t['output_tokens'],'\nReasoning:',t['reasoning'],'\nFinal:',t['answer'])
   print('Inspect arithmetic 8 then 9 and function update manually; no returned code is executed and no model correctness is assumed.')
   PYVERIFY

   Expected result: Unauthenticated access returns 401. Authenticated health and two complete raw-EOS, closed-thinking turns pass exact token-budget checks. Human inspection determines actual task quality; code is not executed.

Performance evidence

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

Limits and troubleshooting

• Only exact source and metadata plus stdlib fake/AST checks were executed. Native environment, weights, kernels, model quality, memory peak and speed remain unmeasured.
• The five-second socket timeout bounds request framing only. Serial generation can run much longer and requires explicit process shutdown; this is not an inference SLA.
• The fixed reasoning/code inputs do not establish unrestricted tool, production, public API or whole-family coverage. Required guide totals remain null.

Asset, native import, GPU, extension or loading failure

Stop and retain diagnostics without secret values. Do not bypass checks, omit experts, change source math, substitute another checkpoint or enable an approximate head. A different platform or backend requires its own qualification.

Length exhaustion, missing EOS or malformed thinking/control output

Treat the result as incomplete. Preserve actual output and token counts; do not strip native control fragments, fabricate a final answer or execute returned code.

Out of memory or client timeout

Stop the foreground process and verify endpoint closure. Native compute is not cancelled by the socket or client timeout. Record actual failure; no fit or latency was measured.

Stop the model

1. 1. Stop and verify endpoint closure

   Press Ctrl-C in the foreground service terminal. Then use this second-terminal check. A stopped service ends native work, not cloud resource billing.

   CommandCopy command

   set -eu
   cd "$HOME/maple-preview-aws-p5-4xlarge-bf16-native"
   .venv/bin/python - <<'PYSTOP'
   import socket
   with socket.socket()as s:
    s.settimeout(3);assert s.connect_ex(('127.0.0.1',9008))!=0
   print('Private service endpoint is closed.')
   PYSTOP

   Expected result: The native Python process has stopped and the loopback endpoint is closed.

Uninstall

1. 1. Remove only this named reader workspace

   After stopping and saving any needed result/version records, remove this exact workspace and its downloaded native assets/environment. Do not remove shared system drivers or unrelated files.

   CommandCopy command

   set -eu
   python3 - <<'PYCLEAN'
   from pathlib import Path
   import shutil
   p=Path.home()/"maple-preview-aws-p5-4xlarge-bf16-native"
   assert p.name=="maple-preview-aws-p5-4xlarge-bf16-native" and p.parent==Path.home() and not p.is_symlink()
   if p.exists():shutil.rmtree(p)
   PYCLEAN

   Expected result: Only the named private workspace is removed; shared machine software remains separate.

Shut down cloud resources

1. 1. End compute and retained-resource billing

   In EC2 us-east-1 stop for later reuse or terminate this exact instance. Review root-volume Delete on termination; delete unwanted detached EBS volumes, snapshots and retained public or Elastic IP resources. Local NVMe is ephemeral. Check final billing. Stopping Python alone does not stop VM or storage charges.

   Expected result: Compute billing ends only after the actual instance lifecycle transition; retained storage, snapshots and IP resources may still bill.