2024-04-18-wg-agenda.md

WebML WG Teleconference – 18 April 2024 - 14:00-15:00 UTC

See the timezone table ...

San Francisco	Thu, 18 April 2024	07:00	UTC-7 hours
Boston	Thu, 18 April 2024	10:00	UTC-4 hours
London	Thu, 18 April 2024	15:00	UTC+1 hours
Berlin	Thu, 18 April 2024	16:00	UTC+2 hours
Helsinki	Thu, 18 April 2024	17:00	UTC+3 hours
Shanghai	Thu, 18 April 2024	22:00	UTC+8 hours
Tokyo	Thu, 18 April 2024	23:00	UTC+9 hours
UTC	Thu, 18 April 2024	14:00 UTC

Other locations: https://www.timeanddate.com/worldclock/fixedtime.html?iso=20240418T14

Logistics

• Chair: Anssi
• Scribe: ? (see howto, volunteers welcome!)
• IRC: irc://irc.w3.org:6667/#webmachinelearning
• IRC web client: https://irc.w3.org/?channels=#webmachinelearning
• Joining instructions: https://lists.w3.org/Archives/Member/internal-webmachinelearning/2023Jun/0000.html
• Minutes: https://www.w3.org/2024/04/18-webmachinelearning-minutes.html

Agenda

ℹ️ WebNN W3C Candidate Recommendation Snapshot published

A new WebNN Candidate Recommendation Snapshot was published 11 April 2024. Since the initial Candidate Recommendation Snapshot published 30 March 2023 the group has gathered further implementation experience and added new operations and data types needed for well-known transformers to support generative AI use cases.

Join us to celebrate this milestone.

• Web Neural Network API W3C Candidate Recommendation Snapshot, 11 April 2024 https://www.w3.org/TR/2024/CR-webnn-20240411/

ℹ️ NPU support discussion (cont'd)

Follow up on the NPU support discussion from our last meeting.

Assess WG readiness to incorporate basic NPU support into the spec. Continue discuss the fallback device proposal and its requirements informed by implementation experience, including backend and driver considerations. Continue discuss the set of ops required to support quantized models targeting NPU.

Please continue provide your feedback via:

• ⨀ WebNN should support NPU and QDQ operations webmachinelearning/webnn#623

Related:

• ⨀ WebNN should support int8 quantized models webmachinelearning/webnn#128
• ⨀ API simplification: context types, context options, createContext() webmachinelearning/webnn#302
• ⨀ Add NPU option for WebNN samples webmachinelearning/webnn-samples#220
• ⛭ Chromium implementation (NPU device type) https://chromium-review.googlesource.com/c/chromium/src/+/5330647

ℹ️ Open issues and PRs

Discuss open issues and review PRs based on your feedback:

• ⨀ All open issues
• ⛙ All open pull requests

A curated shortlist of recently discussed issues below. If any are addressed prior to the meeting, we'll debrief the group on the changes to keep everyone up to date with significant developments. Please refer to the triage guidance for more on our triage practices.

🏷️ Debrief on PRs merged recently

The editors and PR authors will debrief the group on substantive PRs that got merged in the last few weeks, answer question from the group.

• ⛙ Recently merged PRs

🏷️ bug

• Synchronously validate input operands/activations

  • ⨀ webmachinelearning/webnn#572
  • ☰ Any feedback and PRs welcome for the proposed remaining work

• WebIDL definition for constant(start, end, step, type) is missing

  • ⨀ webmachinelearning/webnn#571
  • ☰ Review the proposed WebIDL addition

🏷️ question

• Allow no-op graphs?

  • ⨀ webmachinelearning/webnn#614
  • ☰ Discuss use cases not satisfied by "constant MLBuffer"?

• Can an MLGraphBuilder be reused?

  • ⨀ webmachinelearning/webnn#567
  • ☰ Example code and/or implementation experience wanted

• Specify the operand data type constraints of operation

  • ⨀ webmachinelearning/webnn#283
  • ☰ Review the WIP summary of operand data type constraints for current ops

• Behavior for numeric overflows/underflows/etc in algorithms

  • ⨀ webmachinelearning/webnn#636
  • ☰ Proposed alternatives: do nothing, reference Infra spec, add helpers, add steps to explicitly validate

🏷️ feature request

• Constant sequential filling operation needs output shape parameter
  • ⨀ webmachinelearning/webnn#492
  • ☰ Review new constant() method steps proposal

🏷️ operator specific

• Type of some parameters should match the input data type

  • ⨀ webmachinelearning/webnn#442
  • ☰ Review MLNumber proposal (related #571 #492)

• keepDimensions of reduction ops

  • ⨀ webmachinelearning/webnn#465
  • ☰ Close with an informative note?

🏷️ conventions

No major issues at this time.