2022-03-09-cg-agenda.md

WebML CG Teleconference – 9 March 2022 - 05:00-06:00 UTC+0

See the timezone table ...

San Francisco (U.S.A. - California)	Tue, 8 March 2022	21:00	UTC-8 hours
Boston (U.S.A. - Massachusetts)	Wed, 9 March 2022	00:00	UTC-5 hours
London (United Kingdom - England)	Wed, 9 March 2022	05:00	UTC+0 hours
Berlin (Germany)	Wed, 9 March 2022	06:00	UTC+1 hours
Helsinki (Finland)	Wed, 9 March 2022	07:00	UTC+2 hours
Shanghai (China)	Wed, 9 March 2022	13:00	UTC+8 hours
Tokyo (Japan)	Wed, 9 March 2022	14:00	UTC+9 hours
Sydney (Australia)	Wed, 9 March 2022	16:00	UTC+11 hours
Corresponding UTC (GMT)	Wed, 9 March 2022	05:00 UTC

Other locations: https://www.timeanddate.com/worldclock/fixedtime.html?iso=20220309T05

Logistics

• Chair: Anssi
• Scribe: ? (howto)
• IRC: irc://irc.w3.org:6667/#webmachinelearning
• IRC web client: https://irc.w3.org/?channels=#webmachinelearning
• Call-in details: https://lists.w3.org/Archives/Member/internal-webmachinelearning/2020Apr/0000.html
• Minutes: https://www.w3.org/2022/03/09-webmachinelearning-minutes.html

---

Agenda

🔄 Model Loader API

ℹ️ Versioning

• How to version the API? (supported ops and formats)

ℹ️ Streaming inputs

• Dedicated interface for streaming inputs (e.g. buffered inputs for video, audio)?

ℹ️ Model format

• Prototype vs standard format requirements
• "The Model Loader API needs a standard format supported across browsers and devices for broad interoperability." condition for standards adoption into WebML WG

ℹ️ Support for non-IEEE 754 float point types

• issue: Open question: support for non-IEEE 754 float point types #23 webmachinelearning/model-loader#23

Some accelerators use non-standard float point types (e.g. bfloat16 and TF32) for performance, memory usage optimization. Discuss Model Loader experimentation with these types, open questions:

• Could WebNN and Model Loader API use a common approach for these types?
• How to represent, manipulate these types in JavaScript? Or auto-convert at runtime with some heuristic?
• How to have the two specs complement each other?
• WebNN provides general purpose feature that runs everywhere (write once, run everywhere), Model Loader optionally makes use of accelerator specific features, and could run extremely efficiently based on the model?
• How much information would this give away, privacy risk?

ℹ️ References

• WebML CG Teleconference – 9 Feb 2022 minutes
• WebML CG Teleconference – 12 Jan 2022 minutes
• Model Loader API explainer and early spec draft
• Model Loader API Chromium prototype