Proposal: STC-1.0 Tensor Container for Zero-Copy WebGPU/Wasm Loading and Cryptographic Provenance
Hello WebML team,
As we push larger ML models into the browser via WebGPU and WebNN, fetching and parsing multi-gigabyte models introduces severe memory serialization bottlenecks and main-thread blocking. I am proposing the Sovereign Tensor Architecture (STC-1.0) as a native browser delivery format to eliminate these overheads.
1. Eliminating CPU Byte-Shifting with 64-Byte Alignment
Current formats like SafeTensors and GGUF use arbitrary-length metadata headers, causing the binary payload to land on unaligned memory addresses. STC-1.0 mathematically pads all pre-payload metadata to strict 64-byte physical boundaries. This allows WebAssembly and WebGPU engines to mmap or stream the underlying tensor buffers directly into GPU memory without CPU-side byte-shifting or redundant array cloning.
2. Single HTTP Request Delivery (SIE-1.0)
Web apps currently must orchestrate multiple network requests for config.json, tokenizer.json, and the weights. The Sovereign Inference Engine (SIE-1.0) manifest natively embeds lexical tokenizers and semantic layer routing directly into the STC header, requiring only a single, unified fetch.
3. Browser Sandbox Security & Provenance (STP-1.0)
To prevent malicious payload execution in client browsers, STC integrates a mandatory TRUST block containing an AI-specific C2PA profile, X.509 signature chain, and absolute SHA-256 payload hash. Web runtime engines can cryptographically verify the artifact's integrity and training lineage before allocating memory.
Artifacts:
I would love to discuss how 64-byte aligned containers could optimize the proposed WebNN Model Loader API and whether we could introduce STC as a supported web standard.
Proposal: STC-1.0 Tensor Container for Zero-Copy WebGPU/Wasm Loading and Cryptographic Provenance
Hello WebML team,
As we push larger ML models into the browser via WebGPU and WebNN, fetching and parsing multi-gigabyte models introduces severe memory serialization bottlenecks and main-thread blocking. I am proposing the Sovereign Tensor Architecture (STC-1.0) as a native browser delivery format to eliminate these overheads.
1. Eliminating CPU Byte-Shifting with 64-Byte Alignment
Current formats like SafeTensors and GGUF use arbitrary-length metadata headers, causing the binary payload to land on unaligned memory addresses. STC-1.0 mathematically pads all pre-payload metadata to strict 64-byte physical boundaries. This allows WebAssembly and WebGPU engines to
mmapor stream the underlying tensor buffers directly into GPU memory without CPU-side byte-shifting or redundant array cloning.2. Single HTTP Request Delivery (SIE-1.0)
Web apps currently must orchestrate multiple network requests for
config.json,tokenizer.json, and the weights. The Sovereign Inference Engine (SIE-1.0) manifest natively embeds lexical tokenizers and semantic layer routing directly into the STC header, requiring only a single, unified fetch.3. Browser Sandbox Security & Provenance (STP-1.0)
To prevent malicious payload execution in client browsers, STC integrates a mandatory
TRUSTblock containing an AI-specific C2PA profile, X.509 signature chain, and absolute SHA-256 payload hash. Web runtime engines can cryptographically verify the artifact's integrity and training lineage before allocating memory.Artifacts:
I would love to discuss how 64-byte aligned containers could optimize the proposed WebNN Model Loader API and whether we could introduce STC as a supported web standard.