Meeting minutes
we have a new Zoom meeting setup and due to that you should see a prompt about transcription when you join
… we will not share the transcription publicly, it is only accessible to W3C staff and the chair, and used for reference purposes when reviewing the meeting discussions
<RafaelCintron> Has the meeting URL changed?
Anssi: welcome to Dave Raggett from W3C joining the WebML WG
Web Neural Network API
Repository: webmachinelearning/webnn
TPAC F2F plan for WebNN
WebML WG/CG F2F Agenda - TPAC 2026 (Dublin, Ireland)
<gb> Issue 39 WebML WG/CG F2F Agenda - TPAC 2026 (Dublin, Ireland) (by anssiko) [Agenda+]
Anssi: our plan is as follows
… GMT morning session:
… - current baseline, major milestones since Kobe
… - proposed breaking changes
… - device selection & effective compute policy
… this dedicated session is in response to the customer demand for a mechanism to inform client and cloud execution decisions
… after the first break, LLM enablement
… - dynamic shapes next steps with implementation experience
… - subgraphs and operator set
… - low precision data types and quantization
… after lunch implementation experience and ecosystem
… - demos and discussion on implementation experience
… - customer stories that connect our work to real-world use cases
… - prototyping streaming constants and caching compiled graphs
… as for the WebGPU interop v2, Bryan is on vacation during F2F, so consider deferring this, unless we have WebGPU experts available
… MikeW, anyone from WebGPU-land at TPAC?
MikeW: I might participate remotely
Rafael: I will be at WebGPU F2F next week after TPAC, remotely at TPAC
… after afternoon break
… - security hardening
… - numerics & WPT
… - any horizontal review topics and wrap up
… close with a dinner
Christian: TAG has not talked about WebNN recently so I don't think we need a TAG-WG discussion at TPAC
Anssi: I acknowledge that is a lot to cover in one day
… I will consider your feedback and will prune the agenda accordingly
Anssi: Ehsan, are you involved with privacy reviews, any privacy topics on your list?
Ehsan: nothing on top of my head, will check
Low-precision floating point data types
Anssi: issue #930, explainer PR #938 and spec PR #956
<gb> PR 938 Add low-precision floating point data types explainer (by mklimenko-nv) [Agenda+]
<gb> Issue 930 RFE: Add support for more floating point low-precision ML data types (`bfloat16`, `fp8`, `nvfp4`) (by mtavenrath) [opset] [feature request] [Agenda+]
<gb> PR 956 Add `bfloat16` and `float8e4m3` to the spec (by mklimenko-nv) [Agenda+]
Anssi: explainer PR approved by Dwayne, spec PR review submitted by Dwayne
… I invite the group to review the PRs and thank Mikhail for all the contributions
… I have put low-precision data types next steps to the F2F agenda, so overflow from these PRs will get some F2F time
Mikhail: great feedback from Dwayne, hopefully we can discuss this at F2F
Conv2d groups
Anssi: issue #959
<gb> Issue 959 conv2d: infer groups from input and filter shapes when not specified (by ibelem) [Agenda+]
Anssi: Belem reports MLConv2dOptions.groups IDL defaults to 1
… this means grouped convolutions only validate if groups is passed explicitly
… in these implementations, group count is implied via other information
… the problem is implementations that map TFLite ops 1:1 to WebNN fail validation for grouped convolutions
… proposal to remove the default value for MLConv2dOptions.groups to be able to distinguish between explicitly set group count and the implied group count from other information
Dwayne: not opposed to the change, but is this necessary?
… if TFLite delegate calls into WebNN it could do this computation before reaching WebNN
Ningxin: I think the issue Belem identified is about validation, WebNN requires groups to be divided, internally just to infer the groups count, WebNN does not do strict validation on this
Dwayne: we need to understand this issue a bit more
Anssi: we'll come back to this at F2F if new information emerges
batchNormalization + gelu error tolerance
Anssi: issue #950
<gb> Issue 950 Discrepancy in actual and expected output for subgraph test : batchNormalization options.axis=0 + gelu (by hthadicherla) [Agenda+] [web-platform-tests]
Anssi: implementation feedback from TensorRT-RTX EP with ORT backend
… given a graph with batchNormalization options.axis=0 + gelu
… the strictness checks for batchNormalization + gelu fail error tolerance
… also other tested backends fail this strictness check, ORT CPU, TFLite, LiteRT GPU
Dwayne: this would need to be relaxed some, question whether this combination is interesting, Bruce added fusion tests but if we were to relax this we'd need to set this to at least tolerance of gelu
Anssi: spec would remain as is and this only suggests WPT changes?
Dwayne: correct
Full-sequence output validation
<gb> Issue 949 Validate LSTM/GRU full-sequence output regardless of the `returnSequence` flag (by shiyi9801) [security-tracker]
<gb> PR 955 Enforce validation when creating an MLOperandDescriptor (by shiyi9801)
Anssi: we have a resolution for this issue:
https://
Anssi: PR submitted by Shiyi reviewed by Ningxin, any additional feedback or comments?
… we can merge this PR once all feedback is addressed.
Ningxin: the PR fixes output validation issue, later in review found the spec lacks output operand validation for all operators, the PR has wider scope than it initially had, but this generalization is appropriate and we want to get this done right for all affected ops
Anssi: two LGTMs, ready to land
Accuracy for reductions
<gb> Issue 947 reduceLogSumExp reference decomposition contradicts conformance test; accuracy requirements missing (by mtavenrath) [web-platform-tests]
<gb> PR 957 Add stable decomposition clarification (by mklimenko-nv)
Anssi: similarly, we have a resolution for this:
https://
Anssi: and spec PR is reviewed by Dwayne with minor comments, Ningxin's review requested
<ningxin> I'll take a look
Anssi: we can merge also this PR once all feedback is addressed
Dwayne: just wanted to add a minor comment why we want to have this optimization
Mikhail: fair comment from Dwayne, apologies for missing this, will fix this
Next meeting is TPAC F2F
Anssi: due to close proximity of the TPAC F2F meeting, please consider TPAC an opportunity to discuss any pending issues in person
… our bi-weekly meeting on 22 October 2026 is canceled
… looking forward to seeing many of you in Dublin later this month
… safe travels, and please remember to review the F2F agenda that is shaping up, and submit your comments:
WebML WG/CG F2F Agenda - TPAC 2026 (Dublin, Ireland)
<gb> Issue 39 WebML WG/CG F2F Agenda - TPAC 2026 (Dublin, Ireland) (by anssiko) [Agenda+]