CLIP

Exact model and requirements

Artifact:

OpenAI / CLIP ViT-B/32 native checkpoint (representative family variant) · checkpoint openai/CLIP ViT-B/32

Publisher revision:

d05afc436d78f1c48dc0dbf8e5980a9d471f35f6

Artifact digest:

sha256:40d365715913c9da98579312b702a82c18be219cc2a73407c4526f58eba950af

Download size:

0.33 GiB

Memory requirement (unknown):

Unknown — No exact full-process peak for native CLIP ViT-B/32 CPU float32, image/text encoders, preprocessing and wrapper.

System RAM / disk:

Unknown / Unknown

Placement:

single device. One CPU host; no tensor/pipeline sharding. Entire model on CPU; no device_map, GPU or disk offload.

KV cache:

Non-autoregressive paired encoders; no chat/generation K/V cache is configured.

Workload:

77 token context; Unknown; Not applicable; concurrency 1; batch 1; cache unknown; repetitions Unknown.

Tokenizer:

Native CLIP SimpleTokenizer and exact bpe_simple_vocab_16e6.txt.gz · d05afc436d78f1c48dc0dbf8e5980a9d471f35f6

Sampling / reasoning:

No text sampling. Three fixed labels: a diagram, a dog, a cat; fixed publisher CLIP.png. Native resize/center crop to 224 RGB and normalization; image batch 1, text batch 3 padded to 77. No silent truncation. No chat, reasoning, instruction template, text generation or tool execution. Embeddings are normalized before cosine similarity; logits use native learned logit_scale. Relative label scores are not calibrated probabilities or universal truth.

Runtime prerequisites:

Ubuntu 24.04 x86_64; See prerequisites above.

Official artifact · License

Native OpenAI CLIP repository MIT license and OpenAI-owned ViT-B/32 card MIT declaration are archived; retain copyright/license notices. Native model card recommends research and identifies any deployed use as out of scope. This fixed-input private educational experiment does not establish deployment suitability, safety, fairness, calibration or legal clearance. Inputs/labels are English; no automatic decisions or consequential classification.

• Native BPE vocabulary packaged with pinned runtime · 0 GiB · revision d05afc436d78f1c48dc0dbf8e5980a9d471f35f6
• Publisher CLIP.png research diagram fixture · 0 GiB · revision d05afc436d78f1c48dc0dbf8e5980a9d471f35f6

• 353,976,522 bytes is publisher expected checkpoint content length from HEAD metadata, not a downloaded weight-body measurement or full runtime RAM requirement.
• 32 GiB RAM, 40 GiB disk and four threads are author planning allocations, not measured minima. Both encoders, preprocessing, buffers, dependencies and OS add memory/storage.

Install the runtime

1. 1. Create the pinned CPU research environment

   Install hashed official CPU wheels and exact native CLIP source. CLIP dependencies/transitive packages resolve at installation; save pip freeze and stop on pip check errors. This is documented compatibility, not an executed environment.

   CommandCopy command

   sudo apt-get update
   sudo apt-get install -y python3.12 python3.12-venv git curl ca-certificates openssl
   mkdir -p "$HOME/clip-research/models" "$HOME/clip-research/fixture"
   cd "$HOME/clip-research"
   python3.12 -m venv .venv
   .venv/bin/python -m pip install "https://download.pytorch.org/whl/cpu/torch-2.6.0%2Bcpu-cp312-cp312-linux_x86_64.whl#sha256=59e78aa0c690f70734e42670036d6b541930b8eabbaa18d94e090abf14cc4d91" "https://download-r2.pytorch.org/whl/cpu/torchvision-0.21.0%2Bcpu-cp312-cp312-linux_x86_64.whl#sha256=d6874431e678ba107b60a83f255c33f3755f06bad587b1b919aa514ec325dcd8"
   git clone https://github.com/openai/CLIP.git native-clip
   git -C native-clip checkout --detach d05afc436d78f1c48dc0dbf8e5980a9d471f35f6
   .venv/bin/python -m pip install ./native-clip
   .venv/bin/python -m pip check
   .venv/bin/python -m pip freeze > resolved-environment.txt

   Expected result: Exact checkout and CPU wheel hashes resolve; pip check passes and actual environment is recorded. No GPU driver required.

Download the model

1. 1. Verify checkpoint, native tokenizer vocabulary and fixed fixture

   Only the reader following this guide downloads the weight body. The publisher URL contains the expected SHA-256 and archived HEAD reports 353,976,522 bytes; this author did not download it. Verify every selected input before load.

   CommandCopy command

   cd "$HOME/clip-research"
   curl --fail --location --output models/ViT-B-32.pt 'https://openaipublic.azureedge.net/clip/models/40d365715913c9da98579312b702a82c18be219cc2a73407c4526f58eba950af/ViT-B-32.pt'
   curl --fail --location --output fixture/CLIP.png 'https://raw.githubusercontent.com/openai/CLIP/d05afc436d78f1c48dc0dbf8e5980a9d471f35f6/CLIP.png'
   sha256sum --check <<'SHA256'
   40d365715913c9da98579312b702a82c18be219cc2a73407c4526f58eba950af  models/ViT-B-32.pt
   308a3ca4503f1c7a07803916c369d78c4ef501e5ab7fc727da9b5e1d2f9ec85b  fixture/CLIP.png
   924691ac288e54409236115652ad4aa250f48203de50a9e4722a6ecd48d6804a  native-clip/clip/bpe_simple_vocab_16e6.txt.gz
   SHA256

   Expected result: All three expected hashes report OK. Stop on mismatch. Only trusted publisher checkpoint is used; no arbitrary PyTorch file input.

Configure

1. 1. Create a private research key

   Create a local service key only when absent; preserve existing access and do not print or share it. This is not a provider API key.

   CommandCopy command

   cd "$HOME/clip-research"
   umask 077
   if [ ! -e .api-key ]; then openssl rand -hex 32 > .api-key; fi
   test -s .api-key || { echo "Restore empty key before continuing"; exit 1; }
   chmod 600 .api-key

   Expected result: Private nonempty .api-key exists; existing key preserved.

2. 2. Create the fixed-input research harness

   Reader-authored serial HTTPServer binds loopback and accepts only a fixed fixture identifier. It provides no production service guarantees. It loads the exact local image/model and fixed labels; it does not fetch URLs or accept uploads.

   CommandCopy command

   cd "$HOME/clip-research"
   cat > clip_research.py <<'PYSERVER'
   """OpenTools reader-authored educational demonstrator; not publisher server."""
   import hmac,json,os
   from http.server import BaseHTTPRequestHandler,HTTPServer
   
   def handler_factory(generate,key):
       class Handler(BaseHTTPRequestHandler):
           protocol_version='HTTP/1.0'
           def setup(self):
               super().setup(); self.connection.settimeout(10)
           def log_message(self,*args): pass
           def reply(self,status,data):
               body=json.dumps(data,ensure_ascii=False).encode()
               self.send_response(status); self.send_header('Content-Type','application/json; charset=utf-8')
               self.send_header('Content-Length',str(len(body))); self.end_headers(); self.wfile.write(body)
           def authorized(self):
               value=self.headers.get('Authorization','').encode()
               return hmac.compare_digest(value,('Bearer '+key).encode())
           def do_GET(self):
               if not self.authorized(): return self.reply(401,{'error':'unauthorized'})
               if self.path!='/health': return self.reply(404,{'error':'not_found'})
               self.reply(200,{'status':'ready'})
           def do_POST(self):
               if not self.authorized(): return self.reply(401,{'error':'unauthorized'})
               if self.path!='/v1/clip-research': return self.reply(404,{'error':'not_found'})
               if self.headers.get('Transfer-Encoding') is not None: return self.reply(400,{'error':'transfer_encoding_unsupported'})
               if self.headers.get_content_type()!='application/json': return self.reply(415,{'error':'json_required'})
               lengths=self.headers.get_all('Content-Length',[])
               if len(lengths)!=1 or not lengths[0] or not all('0'<=c<='9' for c in lengths[0]): return self.reply(411,{'error':'content_length_required'})
               if len(lengths[0])>4: return self.reply(413,{'error':'body_limit'})
               n=int(lengths[0])
               if not 1<=n<=1024: return self.reply(413,{'error':'body_limit'})
               try:
                   raw=self.rfile.read(n)
                   if len(raw)!=n: return self.reply(400,{'error':'incomplete_body'})
                   data=json.loads(raw)
               except (ValueError,UnicodeError,TimeoutError): return self.reply(400,{'error':'invalid_json'})
               if type(data)!=dict or data!={'fixture':'publisher-clip-diagram'}:
                   return self.reply(400,{'error':'fixed_fixture_required'})
               try: result=generate()
               except Exception: return self.reply(500,{'error':'generation_failed'})
               self.reply(200,result)
       return Handler
   
   def main():
       import torch,clip
       from PIL import Image
       from clip.simple_tokenizer import SimpleTokenizer
       key=open('.api-key',encoding='ascii').read().strip()
       if len(key)!=64 or any(c not in '0123456789abcdef' for c in key): raise SystemExit('Invalid service key')
       torch.set_num_threads(4)
       model,preprocess=clip.load('models/ViT-B-32.pt',device='cpu',jit=False)
       model.eval()
       labels=['a diagram','a dog','a cat']
       tokenizer=SimpleTokenizer()
       counts=[len(tokenizer.encode(x))+2 for x in labels]
       assert all(2<=n<=77 for n in counts)
       tokens=clip.tokenize(labels,context_length=77,truncate=False)
       assert list(tokens.shape)==[3,77]
       with Image.open('fixture/CLIP.png') as im:
           assert im.size==(2162,762)
           image=preprocess(im.convert('RGB')).unsqueeze(0)
       assert list(image.shape)==[1,3,224,224]
       def generate():
           with torch.inference_mode():
               i=model.encode_image(image).float()
               t=model.encode_text(tokens).float()
               assert list(i.shape)==[1,512] and list(t.shape)==[3,512]
               i=i/i.norm(dim=-1,keepdim=True)
               t=t/t.norm(dim=-1,keepdim=True)
               cosine=i@t.T
               logits=model.logit_scale.exp()*cosine
               assert torch.isfinite(cosine).all() and torch.isfinite(logits).all()
           return {'scope':'private research only; deployed use is outside publisher scope','fixture':'publisher-clip-diagram','labels':labels,'text_tokens_including_bos_eos':counts,'padded_text_shape':[3,77],'image_shape':[1,3,224,224],'image_embedding_shape':[1,512],'text_embedding_shape':[3,512],'cosine':cosine[0].tolist(),'logits':logits[0].tolist(),'highest_label':labels[int(logits[0].argmax())],'calibrated':False}
       HTTPServer(('127.0.0.1',8996),handler_factory(generate,key)).serve_forever()
   if __name__=='__main__': main()
   PYSERVER
   .venv/bin/python -m py_compile clip_research.py

   Expected result: Python syntax check passes without loading the model.

Start the model

1. 1. Start the private fixed-input experiment

   In a foreground terminal, load native eager CPU float32 with four threads. Startup verifies input shapes and native text length per fixed label. One request at a time; no public listener, batching service or automatic decisions.

   CommandCopy command

   cd "$HOME/clip-research"
   .venv/bin/python clip_research.py

   Expected result: After source-supported but unexecuted load, authenticated health can report ready. Stop on missing input, token overflow or preprocessing assertion.

Verify the setup

1. 1. Inspect paired-input research output

   Run from a second terminal on the server. Confirm exact fixed labels, token counts <=77 including start/end, padded text shape 3x77, image 1x3x224x224, embedding shapes 1x512 and 3x512, and three finite cosine/logit values. The source fixture is an architecture diagram containing dog photographs; inspect this taxonomy limitation yourself. Ranking must be internally consistent with reported logits, but no exact ranking or number is promised. A shape/finite smoke check is not in-domain quality, calibration, bias or deployment validation.

   CommandCopy command

   cd "$HOME/clip-research"
   CLIP_RESEARCH_KEY="$(cat .api-key)"
   curl --fail http://127.0.0.1:8996/health -H "Authorization: Bearer $CLIP_RESEARCH_KEY"
   curl --fail http://127.0.0.1:8996/v1/clip-research -H "Authorization: Bearer $CLIP_RESEARCH_KEY" -H 'Content-Type: application/json' -d '{"fixture":"publisher-clip-diagram"}'
   curl -s -o /dev/null -w '%{http_code}\n' http://127.0.0.1:8996/health
   unset CLIP_RESEARCH_KEY

   Expected result: Authorized health ready; research JSON contains shapes, per-label counts, finite cosine/logits and highest_label derived from those logits, calibrated=false. Missing key returns 401. No actual model result is claimed.

Performance evidence

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

Limits and troubleshooting

• Publisher model card places any deployed use outside scope. This private fixed-input research prototype must not be marketed as production suitable, safe, fair or calibrated; no automatic or consequential decisions.
• ViT-B/32 is an explicitly selected representative inside frozen CLIP family. Nine native variants are listed; eight other variants, other backends and broader paired-input workflows remain unqualified.
• 77 is the native text-encoder token limit per candidate English label including start/end, not pixels, full multimodal context or generated output. Three labels form a text batch3 while one image forms image batch1. Output tokens are not applicable.
• Native BPE/vocabulary and 224 RGB resize/center crop/normalization are pinned. A whole diagram with embedded dog photographs is not a representative in-domain image dataset. Human inspection and broader fixed-taxonomy testing are still needed.
• Native image and text features are normalized and compared; learned-scale logits and any taxonomy ranking are relative to labels, not calibrated truth or proof of general image understanding. No exact score/ranking has been measured.
• The expected native checkpoint digest is publisher-provided and HEAD length captured; no weight body or model execution occurred here. Full fit, speed, latency and memory peak are unknown.
• Native jit=False CPU loading converts to float32. No autoregressive K/V cache, generation sampler, chat template or reasoning mode applies.
• Reader-authored HTTPServer is a serial experimental loopback harness with strict fixed JSON, body bound and socket-read timeout. It has no generation deadline, public TLS or production resilience.
• 32 GiB RAM, 40 GiB disk and four threads are planning allocations. Original dated compute/storage and unknown egress/admin/local power remain separate; no hardware performance is inferred.
• System/transitive package dependencies resolve at installation; save actual versions and logs. Python syntax and fake-encoder transport tests do not establish actual model/dependency compatibility or quality.

Hash, shape or tokenizer assertion failure

Stop. Restore exact pinned native source/BPE, publisher checkpoint and fixture; do not bypass native 77-token-per-label error or silently truncate.

401 or request rejected

Use the matching private key and exact fixed fixture JSON. Arbitrary prompts, labels, uploads and remote URLs are intentionally not accepted by this experiment.

Unexpected ranking

Inspect the diagram and its embedded photographs, fixed English labels and preprocessing. Record the result without inventing expected scores or treating the highest label as calibrated truth. Any input/taxonomy change needs new experimental validation, not a claim of another completed guide.

Memory or package load failure

Stop and inspect available memory and exact CPU dependencies. Record errors; full fit is unmeasured. Do not claim successful model execution from transport-only tests.

Stop the model

1. 1. Stop the experiment

   Press Ctrl+C in the server terminal and unset keys in client shells. Confirm loopback port no longer serves.

   CommandCopy command

   unset CLIP_RESEARCH_KEY

   Expected result: Python research harness stops; loopback port 8996 no longer serves.

Uninstall

1. 1. Remove dedicated experiment files

   Preserve wanted logs/results first and confirm the dedicated path. Shared system packages may remain.

   CommandCopy command

   rm -r -- "$HOME/clip-research"

   Expected result: Dedicated checkpoint, fixture, native checkout, environment and key removed.

Exact model and requirements

Artifact:

OpenAI / CLIP ViT-B/32 native checkpoint (representative family variant) · checkpoint openai/CLIP ViT-B/32

Publisher revision:

d05afc436d78f1c48dc0dbf8e5980a9d471f35f6

Artifact digest:

sha256:40d365715913c9da98579312b702a82c18be219cc2a73407c4526f58eba950af

Download size:

0.33 GiB

Memory requirement (unknown):

Unknown — No exact full-process peak for native CLIP ViT-B/32 CPU float32, image/text encoders, preprocessing and wrapper.

System RAM / disk:

Unknown / Unknown

Placement:

single device. One CPU host; no tensor/pipeline sharding. Entire model on CPU; no device_map, GPU or disk offload.

KV cache:

Non-autoregressive paired encoders; no chat/generation K/V cache is configured.

Workload:

77 token context; Unknown; Not applicable; concurrency 1; batch 1; cache unknown; repetitions Unknown.

Tokenizer:

Native CLIP SimpleTokenizer and exact bpe_simple_vocab_16e6.txt.gz · d05afc436d78f1c48dc0dbf8e5980a9d471f35f6

Sampling / reasoning:

No text sampling. Three fixed labels: a diagram, a dog, a cat; fixed publisher CLIP.png. Native resize/center crop to 224 RGB and normalization; image batch 1, text batch 3 padded to 77. No silent truncation. No chat, reasoning, instruction template, text generation or tool execution. Embeddings are normalized before cosine similarity; logits use native learned logit_scale. Relative label scores are not calibrated probabilities or universal truth.

Runtime prerequisites:

Ubuntu 24.04 x86_64; See prerequisites above.

Official artifact · License

Native OpenAI CLIP repository MIT license and OpenAI-owned ViT-B/32 card MIT declaration are archived; retain copyright/license notices. Native model card recommends research and identifies any deployed use as out of scope. This fixed-input private educational experiment does not establish deployment suitability, safety, fairness, calibration or legal clearance. Inputs/labels are English; no automatic decisions or consequential classification.

• Native BPE vocabulary packaged with pinned runtime · 0 GiB · revision d05afc436d78f1c48dc0dbf8e5980a9d471f35f6
• Publisher CLIP.png research diagram fixture · 0 GiB · revision d05afc436d78f1c48dc0dbf8e5980a9d471f35f6

• 353,976,522 bytes is publisher expected checkpoint content length from HEAD metadata, not a downloaded weight-body measurement or full runtime RAM requirement.
• 32 GiB RAM, 40 GiB disk and four threads are author planning allocations, not measured minima. Both encoders, preprocessing, buffers, dependencies and OS add memory/storage.

Install the runtime

1. 1. Create the pinned CPU research environment

   Install hashed official CPU wheels and exact native CLIP source. CLIP dependencies/transitive packages resolve at installation; save pip freeze and stop on pip check errors. This is documented compatibility, not an executed environment.

   CommandCopy command

   sudo apt-get update
   sudo apt-get install -y python3.12 python3.12-venv git curl ca-certificates openssl
   mkdir -p "$HOME/clip-research/models" "$HOME/clip-research/fixture"
   cd "$HOME/clip-research"
   python3.12 -m venv .venv
   .venv/bin/python -m pip install "https://download.pytorch.org/whl/cpu/torch-2.6.0%2Bcpu-cp312-cp312-linux_x86_64.whl#sha256=59e78aa0c690f70734e42670036d6b541930b8eabbaa18d94e090abf14cc4d91" "https://download-r2.pytorch.org/whl/cpu/torchvision-0.21.0%2Bcpu-cp312-cp312-linux_x86_64.whl#sha256=d6874431e678ba107b60a83f255c33f3755f06bad587b1b919aa514ec325dcd8"
   git clone https://github.com/openai/CLIP.git native-clip
   git -C native-clip checkout --detach d05afc436d78f1c48dc0dbf8e5980a9d471f35f6
   .venv/bin/python -m pip install ./native-clip
   .venv/bin/python -m pip check
   .venv/bin/python -m pip freeze > resolved-environment.txt

   Expected result: Exact checkout and CPU wheel hashes resolve; pip check passes and actual environment is recorded. No GPU driver required.

Download the model

1. 1. Verify checkpoint, native tokenizer vocabulary and fixed fixture

   Only the reader following this guide downloads the weight body. The publisher URL contains the expected SHA-256 and archived HEAD reports 353,976,522 bytes; this author did not download it. Verify every selected input before load.

   CommandCopy command

   cd "$HOME/clip-research"
   curl --fail --location --output models/ViT-B-32.pt 'https://openaipublic.azureedge.net/clip/models/40d365715913c9da98579312b702a82c18be219cc2a73407c4526f58eba950af/ViT-B-32.pt'
   curl --fail --location --output fixture/CLIP.png 'https://raw.githubusercontent.com/openai/CLIP/d05afc436d78f1c48dc0dbf8e5980a9d471f35f6/CLIP.png'
   sha256sum --check <<'SHA256'
   40d365715913c9da98579312b702a82c18be219cc2a73407c4526f58eba950af  models/ViT-B-32.pt
   308a3ca4503f1c7a07803916c369d78c4ef501e5ab7fc727da9b5e1d2f9ec85b  fixture/CLIP.png
   924691ac288e54409236115652ad4aa250f48203de50a9e4722a6ecd48d6804a  native-clip/clip/bpe_simple_vocab_16e6.txt.gz
   SHA256

   Expected result: All three expected hashes report OK. Stop on mismatch. Only trusted publisher checkpoint is used; no arbitrary PyTorch file input.

Configure

1. 1. Create a private research key

   Create a local service key only when absent; preserve existing access and do not print or share it. This is not a provider API key.

   CommandCopy command

   cd "$HOME/clip-research"
   umask 077
   if [ ! -e .api-key ]; then openssl rand -hex 32 > .api-key; fi
   test -s .api-key || { echo "Restore empty key before continuing"; exit 1; }
   chmod 600 .api-key

   Expected result: Private nonempty .api-key exists; existing key preserved.

2. 2. Create the fixed-input research harness

   Reader-authored serial HTTPServer binds loopback and accepts only a fixed fixture identifier. It provides no production service guarantees. It loads the exact local image/model and fixed labels; it does not fetch URLs or accept uploads.

   CommandCopy command

   cd "$HOME/clip-research"
   cat > clip_research.py <<'PYSERVER'
   """OpenTools reader-authored educational demonstrator; not publisher server."""
   import hmac,json,os
   from http.server import BaseHTTPRequestHandler,HTTPServer
   
   def handler_factory(generate,key):
       class Handler(BaseHTTPRequestHandler):
           protocol_version='HTTP/1.0'
           def setup(self):
               super().setup(); self.connection.settimeout(10)
           def log_message(self,*args): pass
           def reply(self,status,data):
               body=json.dumps(data,ensure_ascii=False).encode()
               self.send_response(status); self.send_header('Content-Type','application/json; charset=utf-8')
               self.send_header('Content-Length',str(len(body))); self.end_headers(); self.wfile.write(body)
           def authorized(self):
               value=self.headers.get('Authorization','').encode()
               return hmac.compare_digest(value,('Bearer '+key).encode())
           def do_GET(self):
               if not self.authorized(): return self.reply(401,{'error':'unauthorized'})
               if self.path!='/health': return self.reply(404,{'error':'not_found'})
               self.reply(200,{'status':'ready'})
           def do_POST(self):
               if not self.authorized(): return self.reply(401,{'error':'unauthorized'})
               if self.path!='/v1/clip-research': return self.reply(404,{'error':'not_found'})
               if self.headers.get('Transfer-Encoding') is not None: return self.reply(400,{'error':'transfer_encoding_unsupported'})
               if self.headers.get_content_type()!='application/json': return self.reply(415,{'error':'json_required'})
               lengths=self.headers.get_all('Content-Length',[])
               if len(lengths)!=1 or not lengths[0] or not all('0'<=c<='9' for c in lengths[0]): return self.reply(411,{'error':'content_length_required'})
               if len(lengths[0])>4: return self.reply(413,{'error':'body_limit'})
               n=int(lengths[0])
               if not 1<=n<=1024: return self.reply(413,{'error':'body_limit'})
               try:
                   raw=self.rfile.read(n)
                   if len(raw)!=n: return self.reply(400,{'error':'incomplete_body'})
                   data=json.loads(raw)
               except (ValueError,UnicodeError,TimeoutError): return self.reply(400,{'error':'invalid_json'})
               if type(data)!=dict or data!={'fixture':'publisher-clip-diagram'}:
                   return self.reply(400,{'error':'fixed_fixture_required'})
               try: result=generate()
               except Exception: return self.reply(500,{'error':'generation_failed'})
               self.reply(200,result)
       return Handler
   
   def main():
       import torch,clip
       from PIL import Image
       from clip.simple_tokenizer import SimpleTokenizer
       key=open('.api-key',encoding='ascii').read().strip()
       if len(key)!=64 or any(c not in '0123456789abcdef' for c in key): raise SystemExit('Invalid service key')
       torch.set_num_threads(4)
       model,preprocess=clip.load('models/ViT-B-32.pt',device='cpu',jit=False)
       model.eval()
       labels=['a diagram','a dog','a cat']
       tokenizer=SimpleTokenizer()
       counts=[len(tokenizer.encode(x))+2 for x in labels]
       assert all(2<=n<=77 for n in counts)
       tokens=clip.tokenize(labels,context_length=77,truncate=False)
       assert list(tokens.shape)==[3,77]
       with Image.open('fixture/CLIP.png') as im:
           assert im.size==(2162,762)
           image=preprocess(im.convert('RGB')).unsqueeze(0)
       assert list(image.shape)==[1,3,224,224]
       def generate():
           with torch.inference_mode():
               i=model.encode_image(image).float()
               t=model.encode_text(tokens).float()
               assert list(i.shape)==[1,512] and list(t.shape)==[3,512]
               i=i/i.norm(dim=-1,keepdim=True)
               t=t/t.norm(dim=-1,keepdim=True)
               cosine=i@t.T
               logits=model.logit_scale.exp()*cosine
               assert torch.isfinite(cosine).all() and torch.isfinite(logits).all()
           return {'scope':'private research only; deployed use is outside publisher scope','fixture':'publisher-clip-diagram','labels':labels,'text_tokens_including_bos_eos':counts,'padded_text_shape':[3,77],'image_shape':[1,3,224,224],'image_embedding_shape':[1,512],'text_embedding_shape':[3,512],'cosine':cosine[0].tolist(),'logits':logits[0].tolist(),'highest_label':labels[int(logits[0].argmax())],'calibrated':False}
       HTTPServer(('127.0.0.1',8996),handler_factory(generate,key)).serve_forever()
   if __name__=='__main__': main()
   PYSERVER
   .venv/bin/python -m py_compile clip_research.py

   Expected result: Python syntax check passes without loading the model.

Start the model

1. 1. Start the private fixed-input experiment

   In a foreground terminal, load native eager CPU float32 with four threads. Startup verifies input shapes and native text length per fixed label. One request at a time; no public listener, batching service or automatic decisions.

   CommandCopy command

   cd "$HOME/clip-research"
   .venv/bin/python clip_research.py

   Expected result: After source-supported but unexecuted load, authenticated health can report ready. Stop on missing input, token overflow or preprocessing assertion.

Verify the setup

1. 1. Inspect paired-input research output

   Run from a second terminal on the server. Confirm exact fixed labels, token counts <=77 including start/end, padded text shape 3x77, image 1x3x224x224, embedding shapes 1x512 and 3x512, and three finite cosine/logit values. The source fixture is an architecture diagram containing dog photographs; inspect this taxonomy limitation yourself. Ranking must be internally consistent with reported logits, but no exact ranking or number is promised. A shape/finite smoke check is not in-domain quality, calibration, bias or deployment validation.

   CommandCopy command

   cd "$HOME/clip-research"
   CLIP_RESEARCH_KEY="$(cat .api-key)"
   curl --fail http://127.0.0.1:8996/health -H "Authorization: Bearer $CLIP_RESEARCH_KEY"
   curl --fail http://127.0.0.1:8996/v1/clip-research -H "Authorization: Bearer $CLIP_RESEARCH_KEY" -H 'Content-Type: application/json' -d '{"fixture":"publisher-clip-diagram"}'
   curl -s -o /dev/null -w '%{http_code}\n' http://127.0.0.1:8996/health
   unset CLIP_RESEARCH_KEY

   Expected result: Authorized health ready; research JSON contains shapes, per-label counts, finite cosine/logits and highest_label derived from those logits, calibrated=false. Missing key returns 401. No actual model result is claimed.

Performance evidence

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

Limits and troubleshooting

• Publisher model card places any deployed use outside scope. This private fixed-input research prototype must not be marketed as production suitable, safe, fair or calibrated; no automatic or consequential decisions.
• ViT-B/32 is an explicitly selected representative inside frozen CLIP family. Nine native variants are listed; eight other variants, other backends and broader paired-input workflows remain unqualified.
• 77 is the native text-encoder token limit per candidate English label including start/end, not pixels, full multimodal context or generated output. Three labels form a text batch3 while one image forms image batch1. Output tokens are not applicable.
• Native BPE/vocabulary and 224 RGB resize/center crop/normalization are pinned. A whole diagram with embedded dog photographs is not a representative in-domain image dataset. Human inspection and broader fixed-taxonomy testing are still needed.
• Native image and text features are normalized and compared; learned-scale logits and any taxonomy ranking are relative to labels, not calibrated truth or proof of general image understanding. No exact score/ranking has been measured.
• The expected native checkpoint digest is publisher-provided and HEAD length captured; no weight body or model execution occurred here. Full fit, speed, latency and memory peak are unknown.
• Native jit=False CPU loading converts to float32. No autoregressive K/V cache, generation sampler, chat template or reasoning mode applies.
• Reader-authored HTTPServer is a serial experimental loopback harness with strict fixed JSON, body bound and socket-read timeout. It has no generation deadline, public TLS or production resilience.
• 32 GiB RAM, 40 GiB disk and four threads are planning allocations. Original dated compute/storage and unknown egress/admin/local power remain separate; no hardware performance is inferred.
• System/transitive package dependencies resolve at installation; save actual versions and logs. Python syntax and fake-encoder transport tests do not establish actual model/dependency compatibility or quality.

Hash, shape or tokenizer assertion failure

Stop. Restore exact pinned native source/BPE, publisher checkpoint and fixture; do not bypass native 77-token-per-label error or silently truncate.

401 or request rejected

Use the matching private key and exact fixed fixture JSON. Arbitrary prompts, labels, uploads and remote URLs are intentionally not accepted by this experiment.

Unexpected ranking

Inspect the diagram and its embedded photographs, fixed English labels and preprocessing. Record the result without inventing expected scores or treating the highest label as calibrated truth. Any input/taxonomy change needs new experimental validation, not a claim of another completed guide.

Memory or package load failure

Stop and inspect available memory and exact CPU dependencies. Record errors; full fit is unmeasured. Do not claim successful model execution from transport-only tests.

Stop the model

1. 1. Stop the experiment

   Press Ctrl+C in the server terminal and unset keys in client shells. Close SSH tunnel; stopping Python does not stop VM billing.

   CommandCopy command

   unset CLIP_RESEARCH_KEY

   Expected result: Python research harness stops; loopback port 8996 no longer serves.

Uninstall

1. 1. Remove dedicated experiment files

   Preserve wanted logs/results first and confirm the dedicated path. Shared system packages may remain.

   CommandCopy command

   rm -r -- "$HOME/clip-research"

   Expected result: Dedicated checkpoint, fixture, native checkout, environment and key removed.

Shut down cloud resources

1. 1. End cloud billing resources

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

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

Exact model and requirements

Artifact:

OpenAI / Native CLIP ViT-L/14@336px; explicit higher-resolution family variant · checkpoint openai/CLIP ViT-L/14@336px

Publisher revision:

d05afc436d78f1c48dc0dbf8e5980a9d471f35f6

Artifact digest:

sha256:3035c92b350959924f9f00213499208652fc7ea050643e8b385c2dac08641f02

Download size:

Unknown

Memory requirement (unknown):

Unknown — Full native CPU model, preprocessing, text/image attention, buffers and process peak have not been measured.

System RAM / disk:

Unknown / Unknown

Placement:

single device. One CPU host; no tensor/pipeline sharding. Entire model on CPU; no device_map, GPU or disk offload.

KV cache:

CLIP paired encoders do not generate autoregressively or retain a chat K/V cache.

Workload:

77 token context; Unknown; Not applicable; concurrency 1; batch 1; cache cold; repetitions Unknown.

Tokenizer:

Native CLIP SimpleTokenizer, pinned packaged BPE; 77 positions per label including BOS/EOS · d05afc436d78f1c48dc0dbf8e5980a9d471f35f6

Sampling / reasoning:

No generation sampler or chat template. Fixed English labels: a red rectangle on a white background; a blue circle on a white background. Two original 336 × 336 RGB images are generated locally by the reader. Native bicubic Resize, CenterCrop, RGB conversion, ToTensor and exact CLIP normalization. Images encoded serially with batch 1; two labels form text batch 2 padded to 77, truncate=False. The native image transformer has 336/14 = 24 patches per side, 576 patch positions plus one learned class position = 577. This visual sequence is distinct from the per-label 77 text context and from the 768-dimensional projected embeddings. Both encoders are complete; no text generation, decoded output, reasoning or tool execution applies. HF publisher config is corroborating architecture metadata only; native loader shape checks bind the reader-selected native checkpoint, with no native/HF byte-equivalence claim.

Runtime prerequisites:

Ubuntu 24.04 x86_64; See prerequisites above.

Official artifact · License

Actual native OpenAI CLIP repository MIT license is archived; retain notices. The same repository lists this exact publisher checkpoint in its native model map and describes its April 2022 release. No separate HF checkpoint grant or byte equality is inferred. Publisher model card places any deployed use, commercial or not, outside scope and requires thorough in-domain fixed-taxonomy testing even for constrained non-deployed image search. Private/cloud use is not an exemption. This documented educational reproducer supplies no deployment suitability, bias/safety assessment, calibrated accuracy or legal clearance; no surveillance, face recognition, consequential or automatic decisions.

• Native BPE vocabulary packaged with pinned runtime · 0 GiB · revision d05afc436d78f1c48dc0dbf8e5980a9d471f35f6

• Architecture accounting gives 427,944,193 scalar parameters and 1,711,776,772 bytes if all are FP32; this source-derived parameter-storage estimate is not observed checkpoint bytes, total resident RAM, peak or fit. Native reader architecture checks must pass.
• 32 GiB RAM / 40 GiB disk / four threads are reused planning allocations, not measured minima. Native JIT and eager state may coexist during loading; tokenizer, two image tensors, dependency environment, buffers and OS add unknown overhead.
• The native serialized checkpoint size is unknown; no HEAD or body was requested. Separate HF serialization metadata is not substituted for this native artifact.

Install the runtime

1. 1. Create the pinned CPU research environment

   Install hashed official CPU wheels and exact native CLIP source. CLIP dependencies/transitive packages resolve at installation; save pip freeze and stop on pip check errors. This is documented compatibility, not an executed environment. The source package uses the same accepted CPU stack as the B32 guide; changing native architecture is verified by explicit reader loading/shape guards, not inferred execution.

   CommandCopy command

   set -e
   sudo apt-get update
   sudo apt-get install -y python3.12 python3.12-venv git curl ca-certificates openssl
   mkdir -p "$HOME/clip336-research/models" "$HOME/clip336-research/fixture"
   cd "$HOME/clip336-research"
   python3.12 -m venv .venv
   .venv/bin/python -m pip install "https://download.pytorch.org/whl/cpu/torch-2.6.0%2Bcpu-cp312-cp312-linux_x86_64.whl#sha256=59e78aa0c690f70734e42670036d6b541930b8eabbaa18d94e090abf14cc4d91" "https://download-r2.pytorch.org/whl/cpu/torchvision-0.21.0%2Bcpu-cp312-cp312-linux_x86_64.whl#sha256=d6874431e678ba107b60a83f255c33f3755f06bad587b1b919aa514ec325dcd8"
   git clone https://github.com/openai/CLIP.git native-clip
   git -C native-clip checkout --detach d05afc436d78f1c48dc0dbf8e5980a9d471f35f6
   .venv/bin/python -m pip install ./native-clip
   .venv/bin/python -m pip check
   .venv/bin/python -m pip freeze > resolved-environment.txt

   Expected result: Exact checkout and CPU wheel hashes resolve; pip check passes and actual environment is recorded. No GPU driver required.

Download the model

1. 1. Verify the exact native 336px checkpoint and packaged vocabulary

   Only the reader performs this download. The expected digest is literally the parent path of the pinned native OpenAI URL. Exact native file size remains unknown here; no weight body, HEAD or range was requested. Verify complete bytes before native import/load; do not use a similarly named HF or 224px artifact.

   CommandCopy command

   set -e
   cd "$HOME/clip336-research"
   curl --fail --location --output models/ViT-L-14-336px.pt 'https://openaipublic.azureedge.net/clip/models/3035c92b350959924f9f00213499208652fc7ea050643e8b385c2dac08641f02/ViT-L-14-336px.pt'
   sha256sum --check <<'SHA256'
   3035c92b350959924f9f00213499208652fc7ea050643e8b385c2dac08641f02  models/ViT-L-14-336px.pt
   924691ac288e54409236115652ad4aa250f48203de50a9e4722a6ecd48d6804a  native-clip/clip/bpe_simple_vocab_16e6.txt.gz
   SHA256

   Expected result: Both complete hash checks report OK; stop on mismatch. The reader-generated fixtures are created later, with no remote media download or invented publisher fixture digest.

Configure

1. 1. Create a private research key

   Create a local service key only when absent; preserve existing access and do not print or share it. This is not a provider API key.

   CommandCopy command

   set -e
   cd "$HOME/clip336-research"
   umask 077
   if [ ! -e .api-key ]; then openssl rand -hex 32 > .api-key; fi
   test -s .api-key || { echo "Restore empty key before continuing"; exit 1; }
   chmod 600 .api-key

   Expected result: Private nonempty .api-key exists; existing key preserved.

2. 2. Create the native paired-encoder adapter

   Reader-authored code verifies the trusted native checkpoint digest before native imports, loads the exact local file with native jit=False CPU handling and checks full architecture/dtype/finite parameters. It creates two original RGB geometry fixtures and validates native preprocessing/text lengths. This is not publisher server code.

   CommandCopy command

   set -e
   cd "$HOME/clip336-research"
   cat > clip_research.py <<'PYCORE'
   """Reader-authored CLIP native336 fixed paired-image/text research; not publisher server."""
   import hashlib,hmac,json,math,pathlib
   from http.server import BaseHTTPRequestHandler,HTTPServer
   ROOT=pathlib.Path(__file__).resolve().parent
   PORT=8998
   MODEL_SHA='3035c92b350959924f9f00213499208652fc7ea050643e8b385c2dac08641f02'
   LABELS=['a red rectangle on a white background','a blue circle on a white background']
   
   def validate_result(result):
       if type(result)is not dict or set(result)!={'cosine','logits','token_counts'}:raise ValueError('result_schema')
       counts=result['token_counts']
       if type(counts)is not list or len(counts)!=2 or any(type(x)is not int or not 2<=x<=77 for x in counts):raise ValueError('token_counts')
       for name in ['cosine','logits']:
           matrix=result[name]
           if type(matrix)is not list or len(matrix)!=2:raise ValueError('matrix_shape')
           for row in matrix:
               if type(row)is not list or len(row)!=2 or any(type(x)not in(int,float)or not math.isfinite(x)for x in row):raise ValueError('nonfinite_or_shape')
               if name=='cosine'and any(not -1.00001<=x<=1.00001 for x in row):raise ValueError('cosine_range')
       return {'fixture':'author-shapes-v1','variant':'ViT-L/14@336px','labels':LABELS,'fixture_names':['red-rectangle','blue-circle'],'text_tokens_including_bos_eos':counts,'padded_text_shape':[2,77],'image_shape_per_call':[1,3,336,336],'image_embedding_shape_per_call':[1,768],'text_embedding_shape':[2,768],'vision_positions_per_image':577,'cosine':result['cosine'],'logits':result['logits'],'highest_labels':[LABELS[max(range(2),key=lambda i:row[i])]for row in result['logits']],'calibrated':False,'generated_text_tokens':None}
   
   def make_encoder():
       artifact=ROOT/'models/ViT-L-14-336px.pt'
       h=hashlib.sha256()
       with artifact.open('rb')as f:
           for block in iter(lambda:f.read(1048576),b''):h.update(block)
       if h.hexdigest()!=MODEL_SHA:raise RuntimeError('checkpoint_hash')
       # Native imports are only in the reader-run path, after trusted artifact verification.
       import torch,clip
       from PIL import Image,ImageDraw
       from clip.simple_tokenizer import SimpleTokenizer
       torch.set_num_threads(4)
       model,preprocess=clip.load(str(artifact),device='cpu',jit=False)
       model=model.float().eval()
       if model.visual.input_resolution!=336 or tuple(model.visual.conv1.kernel_size)!=(14,14):raise RuntimeError('variant_dimensions')
       if list(model.visual.positional_embedding.shape)!=[577,1024]or len(model.visual.transformer.resblocks)!=24:raise RuntimeError('vision_architecture')
       if model.context_length!=77 or list(model.text_projection.shape)!=[768,768]:raise RuntimeError('text_architecture')
       if any(p.device.type!='cpu'or p.is_floating_point()and p.dtype!=torch.float32 for p in model.parameters()):raise RuntimeError('cpu_fp32')
       if any(not torch.isfinite(p).all()for p in model.parameters()):raise RuntimeError('nonfinite_weights')
       tokenizer=SimpleTokenizer();counts=[len(tokenizer.encode(x))+2 for x in LABELS]
       tokens=clip.tokenize(LABELS,context_length=77,truncate=False)
       if list(tokens.shape)!=[2,77]or any(not 2<=n<=77 for n in counts):raise RuntimeError('text_budget')
       folder=ROOT/'fixture';folder.mkdir(exist_ok=True);images=[]
       # Original RGB images made by the reader; no downloaded media or external rights.
       for name in ['red-rectangle','blue-circle']:
           im=Image.new('RGB',(336,336),(255,255,255));draw=ImageDraw.Draw(im)
           if name=='red-rectangle':draw.rectangle((56,104,280,232),fill=(220,30,30))
           else:draw.ellipse((72,72,264,264),fill=(30,70,220))
           im.save(folder/(name+'.png'))
           image=preprocess(im).unsqueeze(0)
           if list(image.shape)!=[1,3,336,336]or not torch.isfinite(image).all():raise RuntimeError('preprocess')
           images.append(image)
       def encode():
           with torch.inference_mode():
               t=model.encode_text(tokens).float()
               if list(t.shape)!=[2,768]or not torch.isfinite(t).all()or (t.norm(dim=-1)<=0).any():raise RuntimeError('text_features')
               t=t/t.norm(dim=-1,keepdim=True);cosine=[];logits=[]
               for image in images:
                   i=model.encode_image(image).float()
                   if list(i.shape)!=[1,768]or not torch.isfinite(i).all()or (i.norm(dim=-1)<=0).any():raise RuntimeError('image_features')
                   i=i/i.norm(dim=-1,keepdim=True);c=i@t.T;l=model.logit_scale.exp()*c
                   if not torch.isfinite(c).all()or not torch.isfinite(l).all():raise RuntimeError('similarity')
                   cosine.append(c[0].tolist());logits.append(l[0].tolist())
           return {'cosine':cosine,'logits':logits,'token_counts':counts}
       return encode
   
   PYCORE

   Expected result: Core adapter source written. No model load, tokenizer execution or fixture generation occurs until the reader starts it.

3. 3. Complete and verify the private transport

   Append the strict serial loopback transport, then verify the exact complete wrapper. Only the fixed fixture identifier is accepted; inputs cannot request files, URLs, labels or model selections.

   CommandCopy command

   set -e
   cd "$HOME/clip336-research"
   cat >> clip_research.py <<'PYHTTP'
   def create_server(encoder, key, port=PORT):
       if not isinstance(key, str) or len(key) != 64 or any(c not in '0123456789abcdef' for c in key):
           raise ValueError('key_format')
   
       class Handler(BaseHTTPRequestHandler):
           def setup(self):
               super().setup()
               self.connection.settimeout(5)
   
           def log_message(self, *_):
               pass  # Do not log Bearer keys or request bodies.
   
           def reply(self, status, obj):
               data = json.dumps(obj, allow_nan=False, separators=(',', ':')).encode()
               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 authorized(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.authorized():
                   return self.reply(401, {'error': 'unauthorized'})
               if self.path != '/health':
                   return self.reply(404, {'error': 'not_found'})
               return self.reply(200, {'ready': True, 'text_context_per_label': 77,
                                       'generated_text_tokens': None})
   
           def do_POST(self):
               if not self.authorized():
                   return self.reply(401, {'error': 'unauthorized'})
               if self.path != '/v1/clip-research':
                   return self.reply(404, {'error': 'not_found'})
               if self.headers.get('Transfer-Encoding') is not None:
                   return self.reply(400, {'error': 'transfer_encoding_unsupported'})
               lengths = self.headers.get_all('Content-Length', [])
               if len(lengths) != 1:
                   return self.reply(411, {'error': 'content_length_required'})
               raw = lengths[0]
               if not raw or len(raw) > 4 or any(c not in '0123456789' for c in raw):
                   return self.reply(413, {'error': 'invalid_or_excessive_length'})
               n = int(raw)
               if n < 1 or n > 256:
                   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(n)
               except (TimeoutError, OSError):
                   return self.reply(408, {'error': 'body_timeout'})
               if len(body) != n:
                   return self.reply(400, {'error': 'incomplete_body'})
               try:
                   def unique(pairs):
                       obj = {}
                       for k, v in pairs:
                           if k in obj:
                               raise ValueError('duplicate_key')
                           obj[k] = v
                       return obj
                   request = json.loads(body.decode('utf-8'), object_pairs_hook=unique,
                                        parse_constant=lambda _: (_ for _ in ()).throw(ValueError('constant')))
               except (ValueError, UnicodeError, RecursionError):
                   return self.reply(400, {'error': 'invalid_json'})
               if type(request) is not dict or request != {'fixture': 'author-shapes-v1'}:
                   return self.reply(400, {'error': 'fixed_fixture_required'})
               try:
                   result = validate_result(encoder())
               except Exception:
                   return self.reply(500, {'error': 'predictor_or_mask_failure'})
               return self.reply(200, result)
   
       return HTTPServer(('127.0.0.1', port), Handler)
   
   
   if __name__ == '__main__':
       secret = (ROOT / '.api-key').read_text().strip()
       encoder = make_encoder()
       with create_server(encoder, secret) as server:
           server.serve_forever()
   PYHTTP
   .venv/bin/python -m py_compile clip_research.py
   sha256sum --check <<'SHA256'
   806f8f79957e3fed222aa271651c12a473be3b3e92aa70daaa573c399c8fc3bd  clip_research.py
   SHA256

   Expected result: Python syntax and exact whole-wrapper hash verify without native execution.

Start the model

1. 1. Start the bounded research reproducer

   Run from the dedicated workspace. The native loader rebuilds the complete checkpoint and loads state_dict with strict default checking, then converts to CPU float32/eval. Startup requires vision resolution 336, patch 14, 577 visual positions, 24 vision blocks and 768 projections; unexpected variants stop. Two images are encoded serially, one request at a time.

   CommandCopy command

   set -e
   cd "$HOME/clip336-research"
   .venv/bin/python clip_research.py

   Expected result: Only after the reader completes the unexecuted startup checks does loopback 127.0.0.1:8998 serve health. Actual fit, loading success and native results are not claimed.

Verify the setup

1. 1. Inspect finite paired similarities against the actual original images

   Request the fixed experiment from a second server terminal. Check exact labels/fixtures, 77-position padded text, per-label counts including BOS/EOS, serial image shape 1 × 3 × 336 × 336, 768-dimensional projections and two-by-two finite matrices. Open both generated images and inspect whether matched descriptions rank above mismatched descriptions. Record failures and matrix values without requiring a promised ordering or inventing an accuracy result. Geometry is a deliberately simple controlled RGB input, not a natural-image dataset or sufficient in-domain/bias validation.

   CommandCopy command

   set -e
   cd "$HOME/clip336-research"
   CLIP_RESEARCH_KEY="$(cat .api-key)"
   curl --fail http://127.0.0.1:8998/health -H "Authorization: Bearer $CLIP_RESEARCH_KEY"
   curl --fail http://127.0.0.1:8998/v1/clip-research -H "Authorization: Bearer $CLIP_RESEARCH_KEY" -H 'Content-Type: application/json' -d '{"fixture":"author-shapes-v1"}' -o result.json
   curl -s -o /dev/null -w '%{http_code}\n' http://127.0.0.1:8998/health
   unset CLIP_RESEARCH_KEY
   .venv/bin/python - <<'PYVERIFY'
   import json,math
   d=json.load(open('result.json'));labels=['a red rectangle on a white background','a blue circle on a white background']
   assert d['labels']==labels and d['fixture_names']==['red-rectangle','blue-circle']
   assert d['padded_text_shape']==[2,77] and d['image_shape_per_call']==[1,3,336,336]
   assert d['text_embedding_shape']==[2,768] and d['image_embedding_shape_per_call']==[1,768]
   assert d['vision_positions_per_image']==577 and d['generated_text_tokens']is None and d['calibrated']is False
   assert len(d['text_tokens_including_bos_eos'])==2 and all(type(x)is int and 2<=x<=77 for x in d['text_tokens_including_bos_eos'])
   for name in ['cosine','logits']:
    assert len(d[name])==2 and all(len(r)==2 and all(type(x)in(int,float)and math.isfinite(x)for x in r)for r in d[name])
   assert d['highest_labels']==[labels[max(range(2),key=lambda i:r[i])]for r in d['logits']]
   print('Open fixture/red-rectangle.png and fixture/blue-circle.png. Record matched versus mismatched similarities; no ordering or quality guarantee.')
   PYVERIFY

   Expected result: Authorized health and research JSON follow the exact schema; missing authorization returns 401. Structural checks pass only after actual reader execution. Human inspection records rankings and mistakes, without calibration, general accuracy or deployment approval.

Performance evidence

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

Limits and troubleshooting

• Native CLIP model card places any deployed use, commercial or not, outside scope; private/cloud location is not an exemption. The educational fixed-input protocol reproducer does not grant deployment suitability. Non-deployed use still needs thorough fixed-taxonomy in-domain testing; no automatic, consequential, surveillance or face-recognition use.
• ViT-L/14@336px is a distinct higher-resolution native variant in frozen generic CLIP family. Native expected hash comes from the exact pinned runtime URL; no B32, 224px or HF byte equivalence is claimed. Exact native serialized size is unknown without prohibited body/HEAD/range retrieval.
• The full model includes a 24-block width-1024 visual encoder and 12-block width-768 text encoder, projected to 768 dimensions. The 336px/14px patch grid gives 576 patches plus one class position. This differs from the B32 guide; no sibling fit, speed or quality transfer.
• 77 is the native per-label text length including BOS/EOS. It is separate from the 577 visual sequence and 768 embedding dimension; no text generation, chat template, autoregressive K/V cache or decoded output applies.
• Both original RGB geometry images are generated by the reader; no media was downloaded/generated by this author. Their descriptive labels provide a controlled matched/mismatched comparison, not a representative natural-image dataset. No ranking or score is promised.
• Embeddings are checked finite/nonzero before normalization; cosine/logits must be finite and correctly shaped. Learned-scale logits and highest-label ranking are relative to two labels, not calibrated probabilities, semantic truth or general recognition accuracy.
• 32 GiB RAM, 40 GiB disk and four threads are chosen allocations, not minimum requirements or proven fit. Architecture-derived FP32 parameter storage is about 1.71 GB; load-time JIT/eager copies, attention, buffers, environment and OS add unmeasured overhead.
• The native trusted checkpoint loader first tries torch.jit.load and has a state-dict fallback. Never load arbitrary PyTorch files or relax state-key checking; complete publisher hash verification is mandatory before native import/load.
• Strict private serial HTTPServer bounds fixed request body, socket read, labels, image count and output matrices. It has no native inference deadline, public TLS or multiuser resilience. A private VM endpoint remains subject to the publisher research-use qualification.
• Exact direct CPU wheels/native source are pinned as in the accepted stack; transitive/system dependencies resolve at reader install and must be saved with pip freeze. No model import, native tokenizer, preprocessing, load or inference was executed; fake tests only validate transport/output boundaries.
• Other native variants, arbitrary image search, larger taxonomies, alternative backends/OS and full family scope remain unqualified. This pair is not a final guide denominator.

Hash, architecture, dtype or state-key failure

Stop. Restore the exact native 336px checkpoint/source/BPE and CPU environment. Do not substitute the 224px variant, HF weights or smaller encoders, relax strict state checks or silently truncate text.

Wrong or tied relative rankings

Open both original fixture images and inspect the two fixed descriptions. Record cosine/logit matrices and matched versus mismatched outcomes honestly. No exact order, accuracy, calibration or natural-image generalization is promised.

401 or rejected request

Use the stable private key and exact author-shapes-v1 fixture JSON. Arbitrary file/URL inputs, uploads, custom labels and model changes are rejected.

Native CPU load or memory failure

Stop and record exact dependency versions, available memory and errors. Allocation is unmeasured; do not infer CPU fit from source-only fake tests or delete model components to force a smaller workflow.

Stop the model

1. 1. Stop the experiment

   Press Ctrl+C in the server terminal and unset keys in client shells. Confirm loopback port no longer serves.

   CommandCopy command

   set -e
   unset CLIP_RESEARCH_KEY

   Expected result: Python research harness stops; loopback port 8998 no longer serves.

Uninstall

1. 1. Remove dedicated experiment files

   Preserve wanted logs/results first and confirm the dedicated path. Shared system packages may remain.

   CommandCopy command

   set -e
   rm -r -- "$HOME/clip336-research"

   Expected result: Dedicated checkpoint, native source, virtualenv, two original fixtures, result JSON, wrapper and private key removed; shared system packages remain.

Exact model and requirements

Artifact:

OpenAI / Native CLIP ViT-L/14@336px; explicit higher-resolution family variant · checkpoint openai/CLIP ViT-L/14@336px

Publisher revision:

d05afc436d78f1c48dc0dbf8e5980a9d471f35f6

Artifact digest:

sha256:3035c92b350959924f9f00213499208652fc7ea050643e8b385c2dac08641f02

Download size:

Unknown

Memory requirement (unknown):

Unknown — Full native CPU model, preprocessing, text/image attention, buffers and process peak have not been measured.

System RAM / disk:

Unknown / Unknown

Placement:

single device. One CPU host; no tensor/pipeline sharding. Entire model on CPU; no device_map, GPU or disk offload.

KV cache:

CLIP paired encoders do not generate autoregressively or retain a chat K/V cache.

Workload:

77 token context; Unknown; Not applicable; concurrency 1; batch 1; cache cold; repetitions Unknown.

Tokenizer:

Native CLIP SimpleTokenizer, pinned packaged BPE; 77 positions per label including BOS/EOS · d05afc436d78f1c48dc0dbf8e5980a9d471f35f6

Sampling / reasoning:

No generation sampler or chat template. Fixed English labels: a red rectangle on a white background; a blue circle on a white background. Two original 336 × 336 RGB images are generated locally by the reader. Native bicubic Resize, CenterCrop, RGB conversion, ToTensor and exact CLIP normalization. Images encoded serially with batch 1; two labels form text batch 2 padded to 77, truncate=False. The native image transformer has 336/14 = 24 patches per side, 576 patch positions plus one learned class position = 577. This visual sequence is distinct from the per-label 77 text context and from the 768-dimensional projected embeddings. Both encoders are complete; no text generation, decoded output, reasoning or tool execution applies. HF publisher config is corroborating architecture metadata only; native loader shape checks bind the reader-selected native checkpoint, with no native/HF byte-equivalence claim.

Runtime prerequisites:

Ubuntu 24.04 x86_64; See prerequisites above.

Official artifact · License

Actual native OpenAI CLIP repository MIT license is archived; retain notices. The same repository lists this exact publisher checkpoint in its native model map and describes its April 2022 release. No separate HF checkpoint grant or byte equality is inferred. Publisher model card places any deployed use, commercial or not, outside scope and requires thorough in-domain fixed-taxonomy testing even for constrained non-deployed image search. Private/cloud use is not an exemption. This documented educational reproducer supplies no deployment suitability, bias/safety assessment, calibrated accuracy or legal clearance; no surveillance, face recognition, consequential or automatic decisions.

• Native BPE vocabulary packaged with pinned runtime · 0 GiB · revision d05afc436d78f1c48dc0dbf8e5980a9d471f35f6

• Architecture accounting gives 427,944,193 scalar parameters and 1,711,776,772 bytes if all are FP32; this source-derived parameter-storage estimate is not observed checkpoint bytes, total resident RAM, peak or fit. Native reader architecture checks must pass.
• 32 GiB RAM / 40 GiB disk / four threads are reused planning allocations, not measured minima. Native JIT and eager state may coexist during loading; tokenizer, two image tensors, dependency environment, buffers and OS add unknown overhead.
• The native serialized checkpoint size is unknown; no HEAD or body was requested. Separate HF serialization metadata is not substituted for this native artifact.

Install the runtime

1. 1. Create the pinned CPU research environment

   Install hashed official CPU wheels and exact native CLIP source. CLIP dependencies/transitive packages resolve at installation; save pip freeze and stop on pip check errors. This is documented compatibility, not an executed environment. The source package uses the same accepted CPU stack as the B32 guide; changing native architecture is verified by explicit reader loading/shape guards, not inferred execution.

   CommandCopy command

   set -e
   sudo apt-get update
   sudo apt-get install -y python3.12 python3.12-venv git curl ca-certificates openssl
   mkdir -p "$HOME/clip336-research/models" "$HOME/clip336-research/fixture"
   cd "$HOME/clip336-research"
   python3.12 -m venv .venv
   .venv/bin/python -m pip install "https://download.pytorch.org/whl/cpu/torch-2.6.0%2Bcpu-cp312-cp312-linux_x86_64.whl#sha256=59e78aa0c690f70734e42670036d6b541930b8eabbaa18d94e090abf14cc4d91" "https://download-r2.pytorch.org/whl/cpu/torchvision-0.21.0%2Bcpu-cp312-cp312-linux_x86_64.whl#sha256=d6874431e678ba107b60a83f255c33f3755f06bad587b1b919aa514ec325dcd8"
   git clone https://github.com/openai/CLIP.git native-clip
   git -C native-clip checkout --detach d05afc436d78f1c48dc0dbf8e5980a9d471f35f6
   .venv/bin/python -m pip install ./native-clip
   .venv/bin/python -m pip check
   .venv/bin/python -m pip freeze > resolved-environment.txt

   Expected result: Exact checkout and CPU wheel hashes resolve; pip check passes and actual environment is recorded. No GPU driver required.

Download the model

1. 1. Verify the exact native 336px checkpoint and packaged vocabulary

   Only the reader performs this download. The expected digest is literally the parent path of the pinned native OpenAI URL. Exact native file size remains unknown here; no weight body, HEAD or range was requested. Verify complete bytes before native import/load; do not use a similarly named HF or 224px artifact.

   CommandCopy command

   set -e
   cd "$HOME/clip336-research"
   curl --fail --location --output models/ViT-L-14-336px.pt 'https://openaipublic.azureedge.net/clip/models/3035c92b350959924f9f00213499208652fc7ea050643e8b385c2dac08641f02/ViT-L-14-336px.pt'
   sha256sum --check <<'SHA256'
   3035c92b350959924f9f00213499208652fc7ea050643e8b385c2dac08641f02  models/ViT-L-14-336px.pt
   924691ac288e54409236115652ad4aa250f48203de50a9e4722a6ecd48d6804a  native-clip/clip/bpe_simple_vocab_16e6.txt.gz
   SHA256

   Expected result: Both complete hash checks report OK; stop on mismatch. The reader-generated fixtures are created later, with no remote media download or invented publisher fixture digest.

Configure

1. 1. Create a private research key

   Create a local service key only when absent; preserve existing access and do not print or share it. This is not a provider API key.

   CommandCopy command

   set -e
   cd "$HOME/clip336-research"
   umask 077
   if [ ! -e .api-key ]; then openssl rand -hex 32 > .api-key; fi
   test -s .api-key || { echo "Restore empty key before continuing"; exit 1; }
   chmod 600 .api-key

   Expected result: Private nonempty .api-key exists; existing key preserved.

2. 2. Create the native paired-encoder adapter

   Reader-authored code verifies the trusted native checkpoint digest before native imports, loads the exact local file with native jit=False CPU handling and checks full architecture/dtype/finite parameters. It creates two original RGB geometry fixtures and validates native preprocessing/text lengths. This is not publisher server code.

   CommandCopy command

   set -e
   cd "$HOME/clip336-research"
   cat > clip_research.py <<'PYCORE'
   """Reader-authored CLIP native336 fixed paired-image/text research; not publisher server."""
   import hashlib,hmac,json,math,pathlib
   from http.server import BaseHTTPRequestHandler,HTTPServer
   ROOT=pathlib.Path(__file__).resolve().parent
   PORT=8998
   MODEL_SHA='3035c92b350959924f9f00213499208652fc7ea050643e8b385c2dac08641f02'
   LABELS=['a red rectangle on a white background','a blue circle on a white background']
   
   def validate_result(result):
       if type(result)is not dict or set(result)!={'cosine','logits','token_counts'}:raise ValueError('result_schema')
       counts=result['token_counts']
       if type(counts)is not list or len(counts)!=2 or any(type(x)is not int or not 2<=x<=77 for x in counts):raise ValueError('token_counts')
       for name in ['cosine','logits']:
           matrix=result[name]
           if type(matrix)is not list or len(matrix)!=2:raise ValueError('matrix_shape')
           for row in matrix:
               if type(row)is not list or len(row)!=2 or any(type(x)not in(int,float)or not math.isfinite(x)for x in row):raise ValueError('nonfinite_or_shape')
               if name=='cosine'and any(not -1.00001<=x<=1.00001 for x in row):raise ValueError('cosine_range')
       return {'fixture':'author-shapes-v1','variant':'ViT-L/14@336px','labels':LABELS,'fixture_names':['red-rectangle','blue-circle'],'text_tokens_including_bos_eos':counts,'padded_text_shape':[2,77],'image_shape_per_call':[1,3,336,336],'image_embedding_shape_per_call':[1,768],'text_embedding_shape':[2,768],'vision_positions_per_image':577,'cosine':result['cosine'],'logits':result['logits'],'highest_labels':[LABELS[max(range(2),key=lambda i:row[i])]for row in result['logits']],'calibrated':False,'generated_text_tokens':None}
   
   def make_encoder():
       artifact=ROOT/'models/ViT-L-14-336px.pt'
       h=hashlib.sha256()
       with artifact.open('rb')as f:
           for block in iter(lambda:f.read(1048576),b''):h.update(block)
       if h.hexdigest()!=MODEL_SHA:raise RuntimeError('checkpoint_hash')
       # Native imports are only in the reader-run path, after trusted artifact verification.
       import torch,clip
       from PIL import Image,ImageDraw
       from clip.simple_tokenizer import SimpleTokenizer
       torch.set_num_threads(4)
       model,preprocess=clip.load(str(artifact),device='cpu',jit=False)
       model=model.float().eval()
       if model.visual.input_resolution!=336 or tuple(model.visual.conv1.kernel_size)!=(14,14):raise RuntimeError('variant_dimensions')
       if list(model.visual.positional_embedding.shape)!=[577,1024]or len(model.visual.transformer.resblocks)!=24:raise RuntimeError('vision_architecture')
       if model.context_length!=77 or list(model.text_projection.shape)!=[768,768]:raise RuntimeError('text_architecture')
       if any(p.device.type!='cpu'or p.is_floating_point()and p.dtype!=torch.float32 for p in model.parameters()):raise RuntimeError('cpu_fp32')
       if any(not torch.isfinite(p).all()for p in model.parameters()):raise RuntimeError('nonfinite_weights')
       tokenizer=SimpleTokenizer();counts=[len(tokenizer.encode(x))+2 for x in LABELS]
       tokens=clip.tokenize(LABELS,context_length=77,truncate=False)
       if list(tokens.shape)!=[2,77]or any(not 2<=n<=77 for n in counts):raise RuntimeError('text_budget')
       folder=ROOT/'fixture';folder.mkdir(exist_ok=True);images=[]
       # Original RGB images made by the reader; no downloaded media or external rights.
       for name in ['red-rectangle','blue-circle']:
           im=Image.new('RGB',(336,336),(255,255,255));draw=ImageDraw.Draw(im)
           if name=='red-rectangle':draw.rectangle((56,104,280,232),fill=(220,30,30))
           else:draw.ellipse((72,72,264,264),fill=(30,70,220))
           im.save(folder/(name+'.png'))
           image=preprocess(im).unsqueeze(0)
           if list(image.shape)!=[1,3,336,336]or not torch.isfinite(image).all():raise RuntimeError('preprocess')
           images.append(image)
       def encode():
           with torch.inference_mode():
               t=model.encode_text(tokens).float()
               if list(t.shape)!=[2,768]or not torch.isfinite(t).all()or (t.norm(dim=-1)<=0).any():raise RuntimeError('text_features')
               t=t/t.norm(dim=-1,keepdim=True);cosine=[];logits=[]
               for image in images:
                   i=model.encode_image(image).float()
                   if list(i.shape)!=[1,768]or not torch.isfinite(i).all()or (i.norm(dim=-1)<=0).any():raise RuntimeError('image_features')
                   i=i/i.norm(dim=-1,keepdim=True);c=i@t.T;l=model.logit_scale.exp()*c
                   if not torch.isfinite(c).all()or not torch.isfinite(l).all():raise RuntimeError('similarity')
                   cosine.append(c[0].tolist());logits.append(l[0].tolist())
           return {'cosine':cosine,'logits':logits,'token_counts':counts}
       return encode
   
   PYCORE

   Expected result: Core adapter source written. No model load, tokenizer execution or fixture generation occurs until the reader starts it.

3. 3. Complete and verify the private transport

   Append the strict serial loopback transport, then verify the exact complete wrapper. Only the fixed fixture identifier is accepted; inputs cannot request files, URLs, labels or model selections.

   CommandCopy command

   set -e
   cd "$HOME/clip336-research"
   cat >> clip_research.py <<'PYHTTP'
   def create_server(encoder, key, port=PORT):
       if not isinstance(key, str) or len(key) != 64 or any(c not in '0123456789abcdef' for c in key):
           raise ValueError('key_format')
   
       class Handler(BaseHTTPRequestHandler):
           def setup(self):
               super().setup()
               self.connection.settimeout(5)
   
           def log_message(self, *_):
               pass  # Do not log Bearer keys or request bodies.
   
           def reply(self, status, obj):
               data = json.dumps(obj, allow_nan=False, separators=(',', ':')).encode()
               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 authorized(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.authorized():
                   return self.reply(401, {'error': 'unauthorized'})
               if self.path != '/health':
                   return self.reply(404, {'error': 'not_found'})
               return self.reply(200, {'ready': True, 'text_context_per_label': 77,
                                       'generated_text_tokens': None})
   
           def do_POST(self):
               if not self.authorized():
                   return self.reply(401, {'error': 'unauthorized'})
               if self.path != '/v1/clip-research':
                   return self.reply(404, {'error': 'not_found'})
               if self.headers.get('Transfer-Encoding') is not None:
                   return self.reply(400, {'error': 'transfer_encoding_unsupported'})
               lengths = self.headers.get_all('Content-Length', [])
               if len(lengths) != 1:
                   return self.reply(411, {'error': 'content_length_required'})
               raw = lengths[0]
               if not raw or len(raw) > 4 or any(c not in '0123456789' for c in raw):
                   return self.reply(413, {'error': 'invalid_or_excessive_length'})
               n = int(raw)
               if n < 1 or n > 256:
                   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(n)
               except (TimeoutError, OSError):
                   return self.reply(408, {'error': 'body_timeout'})
               if len(body) != n:
                   return self.reply(400, {'error': 'incomplete_body'})
               try:
                   def unique(pairs):
                       obj = {}
                       for k, v in pairs:
                           if k in obj:
                               raise ValueError('duplicate_key')
                           obj[k] = v
                       return obj
                   request = json.loads(body.decode('utf-8'), object_pairs_hook=unique,
                                        parse_constant=lambda _: (_ for _ in ()).throw(ValueError('constant')))
               except (ValueError, UnicodeError, RecursionError):
                   return self.reply(400, {'error': 'invalid_json'})
               if type(request) is not dict or request != {'fixture': 'author-shapes-v1'}:
                   return self.reply(400, {'error': 'fixed_fixture_required'})
               try:
                   result = validate_result(encoder())
               except Exception:
                   return self.reply(500, {'error': 'predictor_or_mask_failure'})
               return self.reply(200, result)
   
       return HTTPServer(('127.0.0.1', port), Handler)
   
   
   if __name__ == '__main__':
       secret = (ROOT / '.api-key').read_text().strip()
       encoder = make_encoder()
       with create_server(encoder, secret) as server:
           server.serve_forever()
   PYHTTP
   .venv/bin/python -m py_compile clip_research.py
   sha256sum --check <<'SHA256'
   806f8f79957e3fed222aa271651c12a473be3b3e92aa70daaa573c399c8fc3bd  clip_research.py
   SHA256

   Expected result: Python syntax and exact whole-wrapper hash verify without native execution.

Start the model

1. 1. Start the bounded research reproducer

   Run from the dedicated workspace. The native loader rebuilds the complete checkpoint and loads state_dict with strict default checking, then converts to CPU float32/eval. Startup requires vision resolution 336, patch 14, 577 visual positions, 24 vision blocks and 768 projections; unexpected variants stop. Two images are encoded serially, one request at a time.

   CommandCopy command

   set -e
   cd "$HOME/clip336-research"
   .venv/bin/python clip_research.py

   Expected result: Only after the reader completes the unexecuted startup checks does loopback 127.0.0.1:8998 serve health. Actual fit, loading success and native results are not claimed.

Verify the setup

1. 1. Inspect finite paired similarities against the actual original images

   Request the fixed experiment from a second server terminal. Check exact labels/fixtures, 77-position padded text, per-label counts including BOS/EOS, serial image shape 1 × 3 × 336 × 336, 768-dimensional projections and two-by-two finite matrices. Open both generated images and inspect whether matched descriptions rank above mismatched descriptions. Record failures and matrix values without requiring a promised ordering or inventing an accuracy result. Geometry is a deliberately simple controlled RGB input, not a natural-image dataset or sufficient in-domain/bias validation.

   CommandCopy command

   set -e
   cd "$HOME/clip336-research"
   CLIP_RESEARCH_KEY="$(cat .api-key)"
   curl --fail http://127.0.0.1:8998/health -H "Authorization: Bearer $CLIP_RESEARCH_KEY"
   curl --fail http://127.0.0.1:8998/v1/clip-research -H "Authorization: Bearer $CLIP_RESEARCH_KEY" -H 'Content-Type: application/json' -d '{"fixture":"author-shapes-v1"}' -o result.json
   curl -s -o /dev/null -w '%{http_code}\n' http://127.0.0.1:8998/health
   unset CLIP_RESEARCH_KEY
   .venv/bin/python - <<'PYVERIFY'
   import json,math
   d=json.load(open('result.json'));labels=['a red rectangle on a white background','a blue circle on a white background']
   assert d['labels']==labels and d['fixture_names']==['red-rectangle','blue-circle']
   assert d['padded_text_shape']==[2,77] and d['image_shape_per_call']==[1,3,336,336]
   assert d['text_embedding_shape']==[2,768] and d['image_embedding_shape_per_call']==[1,768]
   assert d['vision_positions_per_image']==577 and d['generated_text_tokens']is None and d['calibrated']is False
   assert len(d['text_tokens_including_bos_eos'])==2 and all(type(x)is int and 2<=x<=77 for x in d['text_tokens_including_bos_eos'])
   for name in ['cosine','logits']:
    assert len(d[name])==2 and all(len(r)==2 and all(type(x)in(int,float)and math.isfinite(x)for x in r)for r in d[name])
   assert d['highest_labels']==[labels[max(range(2),key=lambda i:r[i])]for r in d['logits']]
   print('Open fixture/red-rectangle.png and fixture/blue-circle.png. Record matched versus mismatched similarities; no ordering or quality guarantee.')
   PYVERIFY

   Expected result: Authorized health and research JSON follow the exact schema; missing authorization returns 401. Structural checks pass only after actual reader execution. Human inspection records rankings and mistakes, without calibration, general accuracy or deployment approval.

Performance evidence

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

Limits and troubleshooting

• Native CLIP model card places any deployed use, commercial or not, outside scope; private/cloud location is not an exemption. The educational fixed-input protocol reproducer does not grant deployment suitability. Non-deployed use still needs thorough fixed-taxonomy in-domain testing; no automatic, consequential, surveillance or face-recognition use.
• ViT-L/14@336px is a distinct higher-resolution native variant in frozen generic CLIP family. Native expected hash comes from the exact pinned runtime URL; no B32, 224px or HF byte equivalence is claimed. Exact native serialized size is unknown without prohibited body/HEAD/range retrieval.
• The full model includes a 24-block width-1024 visual encoder and 12-block width-768 text encoder, projected to 768 dimensions. The 336px/14px patch grid gives 576 patches plus one class position. This differs from the B32 guide; no sibling fit, speed or quality transfer.
• 77 is the native per-label text length including BOS/EOS. It is separate from the 577 visual sequence and 768 embedding dimension; no text generation, chat template, autoregressive K/V cache or decoded output applies.
• Both original RGB geometry images are generated by the reader; no media was downloaded/generated by this author. Their descriptive labels provide a controlled matched/mismatched comparison, not a representative natural-image dataset. No ranking or score is promised.
• Embeddings are checked finite/nonzero before normalization; cosine/logits must be finite and correctly shaped. Learned-scale logits and highest-label ranking are relative to two labels, not calibrated probabilities, semantic truth or general recognition accuracy.
• 32 GiB RAM, 40 GiB disk and four threads are chosen allocations, not minimum requirements or proven fit. Architecture-derived FP32 parameter storage is about 1.71 GB; load-time JIT/eager copies, attention, buffers, environment and OS add unmeasured overhead.
• The native trusted checkpoint loader first tries torch.jit.load and has a state-dict fallback. Never load arbitrary PyTorch files or relax state-key checking; complete publisher hash verification is mandatory before native import/load.
• Strict private serial HTTPServer bounds fixed request body, socket read, labels, image count and output matrices. It has no native inference deadline, public TLS or multiuser resilience. A private VM endpoint remains subject to the publisher research-use qualification.
• Exact direct CPU wheels/native source are pinned as in the accepted stack; transitive/system dependencies resolve at reader install and must be saved with pip freeze. No model import, native tokenizer, preprocessing, load or inference was executed; fake tests only validate transport/output boundaries.
• Other native variants, arbitrary image search, larger taxonomies, alternative backends/OS and full family scope remain unqualified. This pair is not a final guide denominator.

Hash, architecture, dtype or state-key failure

Stop. Restore the exact native 336px checkpoint/source/BPE and CPU environment. Do not substitute the 224px variant, HF weights or smaller encoders, relax strict state checks or silently truncate text.

Wrong or tied relative rankings

Open both original fixture images and inspect the two fixed descriptions. Record cosine/logit matrices and matched versus mismatched outcomes honestly. No exact order, accuracy, calibration or natural-image generalization is promised.

401 or rejected request

Use the stable private key and exact author-shapes-v1 fixture JSON. Arbitrary file/URL inputs, uploads, custom labels and model changes are rejected.

Native CPU load or memory failure

Stop and record exact dependency versions, available memory and errors. Allocation is unmeasured; do not infer CPU fit from source-only fake tests or delete model components to force a smaller workflow.

Stop the model

1. 1. Stop the experiment

   Press Ctrl+C in the server terminal and unset keys in client shells. Close SSH tunnel; stopping Python does not stop VM billing.

   CommandCopy command

   set -e
   unset CLIP_RESEARCH_KEY

   Expected result: Python research harness stops; loopback port 8998 no longer serves.

Uninstall

1. 1. Remove dedicated experiment files

   Preserve wanted logs/results first and confirm the dedicated path. Shared system packages may remain.

   CommandCopy command

   set -e
   rm -r -- "$HOME/clip336-research"

   Expected result: Dedicated checkpoint, native source, virtualenv, two original fixtures, result JSON, wrapper and private key removed; shared system packages remain.

Shut down cloud resources

1. 1. End cloud billing resources

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

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