WebML WG Teleconference – 22 May 2025 - 15:00-16:00 UTC

Latest version: https://github.com/webmachinelearning/meetings/blob/main/telcons/2025-05-22-wg-agenda.md


WebML WG Teleconference – 22 May 2025 - 15:00-16:00 UTC
<https://github.com/webmachinelearning/meetings/blob/main/telcons/2025-05-22-wg-agenda.md#webml-wg-teleconference--22-may-2025---1500-1600-utc>
See the timezone table ...
San Francisco   Thu, 22 May 2025        08:00
Boston  Thu, 22 May 2025        11:00
London  Thu, 22 May 2025        16:00
Berlin  Thu, 22 May 2025        17:00
Helsinki        Thu, 22 May 2025        18:00
Shanghai        Thu, 22 May 2025        23:00
Tokyo   Fri, 23 May 2025        00:00
UTC     Thu, 22 May 2025        15:00 UTC

Other locations: https://www.timeanddate.com/worldclock/fixedtime.html?iso=20250522T15


Logistics
<https://github.com/webmachinelearning/meetings/blob/main/telcons/2025-05-22-wg-agenda.md#logistics>

  *   Chair: Anssi
  *   Scribe: Anssi
  *   IRC: irc://irc.w3.org:6667/#webmachinelearning

  *   IRC web client: https://irc.w3.org/?channels=#webmachinelearning

  *   Zoom joining instructions: https://lists.w3.org/Archives/Member/internal-webmachinelearning/2023Jun/0000.html

  *   Minutes: https://www.w3.org/2025/05/22-webmachinelearning-minutes.html


Agenda
<https://github.com/webmachinelearning/meetings/blob/main/telcons/2025-05-22-wg-agenda.md#agenda>
🧪 Incubations
<https://github.com/webmachinelearning/meetings/blob/main/telcons/2025-05-22-wg-agenda.md#-incubations>

Discuss recent WebML Community Group developments to keep the Working Group abreast of incubation progress.

New proposals:

  *   ✨ Web AI for Time Series webmachinelearning/proposals#10<https://github.com/webmachinelearning/proposals/issues/10>
  *   ✨ Local Inference Web extension webmachinelearning/proposals#9<https://github.com/webmachinelearning/proposals/issues/9>

🤝 Google I/O and MS Build 2025 takeaways
<https://github.com/webmachinelearning/meetings/blob/main/telcons/2025-05-22-wg-agenda.md#-google-io-and-ms-build-2025-takeaways>

Discuss any updates relevant to the group.

🏷️ Operator specific issues
<https://github.com/webmachinelearning/meetings/blob/main/telcons/2025-05-22-wg-agenda.md#%EF%B8%8F-operator-specific-issues>

Review and discuss operator specific issues that reduce code complexity and improve maintainability, e.g.:

  *   int64 data type

     *   ⨀ Constraints webmachinelearning/webnn#283<https://github.com/webmachinelearning/webnn/issues/283>
     *   ⨀ Reduce ops webmachinelearning/webnn#694<https://github.com/webmachinelearning/webnn/issues/694>
     *   ⛙ Reduce ops (PR) webmachinelearning/webnn#695<https://github.com/webmachinelearning/webnn/pull/695>
     *   ⨀ Ops that take signed ints webmachinelearning/webnn#845<https://github.com/webmachinelearning/webnn/issues/845>
  *   triangular removal

     *   ⨀ Consider new per-model Trilu op count data webmachinelearning/webnn#768<https://github.com/webmachinelearning/webnn/issues/768>

🏷️ Other issues
<https://github.com/webmachinelearning/meetings/blob/main/telcons/2025-05-22-wg-agenda.md#%EF%B8%8F-other-issues>

Review and discuss other issues that benefit from further input and review, e.g.:

  *   opSupportLimits level of detail for output tensor(s) webmachinelearning/webnn#835<https://github.com/webmachinelearning/webnn/issues/835>
  *   round behavior on Core ML NPU/ANE, emulation path webmachinelearning/webnn#817<https://github.com/webmachinelearning/webnn/issues/817>
  *   isNaN op proposal webmachinelearning/webnn#811<https://github.com/webmachinelearning/webnn/issues/811> (see also 2025-03-27 discussion<https://www.w3.org/2025/03/27-webmachinelearning-minutes.html#b06c>)

ℹ️ Caching mechanism for MLGraph
<https://github.com/webmachinelearning/meetings/blob/main/telcons/2025-05-22-wg-agenda.md#%E2%84%B9%EF%B8%8F-caching-mechanism-for-mlgraph>

Revisit prototype implementation findings to reinvigorate work on the explainer.

  *   ⨀ webmachinelearning/webnn#807<https://github.com/webmachinelearning/webnn/issues/807>
  *   ☰ Explainer: WIP
  *   ⚙️ Initial implementation (Chromium + ORT): shiyi9801/chromium#227<https://github.com/shiyi9801/chromium/pull/227> (usage example<https://github.com/webmachinelearning/webnn-samples/compare/master...shiyi9801:webnn-samples:model_cache>)

ℹ️ Query supported devices
<https://github.com/webmachinelearning/meetings/blob/main/telcons/2025-05-22-wg-agenda.md#%E2%84%B9%EF%B8%8F-query-supported-devices>

Solicit further input on high-value use cases to help reduce the solution space:

  *   ⨀ Before graph compilation webmachinelearning/webnn#815<https://github.com/webmachinelearning/webnn/issues/815>

     *   Considered hints: "prefer CPU", "prefer NPU", "prefer GPU", "maximum performance", "maximum efficiency", "minimum overall power"
  *   ⨀ After graph compilation webmachinelearning/webnn#836<https://github.com/webmachinelearning/webnn/issues/836>

     *   Proposed explicit i) devices-to-graph ii) op-to-device map

ℹ️ Core operator set
<https://github.com/webmachinelearning/meetings/blob/main/telcons/2025-05-22-wg-agenda.md#%E2%84%B9%EF%B8%8F-core-operator-set>

(To be discussed subject to interest/new information:)

Revisit our core operator set effort that aims to identify current primitive gaps by mapping compositional fundamentals to WebNN operators.

Discuss any new information on rounding behavior across backends to understand feasibility for inclusion into the core operator set to help with e.g. quantization decomposition.

  *   ⨀ webmachinelearning/webnn#573<https://github.com/webmachinelearning/webnn/issues/573>
  *   ☰ Machine Learning Operator Mapping - All Raw Operators<https://onedrive.live.com/edit?id=EE82F5C6F06C7371!345450&resid=EE82F5C6F06C7371!345450&ithint=file%2Cxlsx&authkey=!AK8f-RDTleqlLXE&wdo=2&cid=ee82f5c6f06c7371>