15:00:04 RRSAgent has joined #webmachinelearning 15:00:08 logging to https://www.w3.org/2026/10/08-webmachinelearning-irc 15:00:08 RRSAgent, make logs Public 15:00:09 please title this meeting ("meeting: ..."), anssik 15:00:12 Meeting: WebML WG Teleconference – 8 October 2026 15:00:15 Chair: Anssi 15:00:19 Agenda: https://github.com/webmachinelearning/meetings/blob/main/telcons/2026-10-08-wg-agenda.md 15:00:24 Scribe: Anssi 15:00:28 scribeNick: anssik 15:00:34 Present+ Anssi_Kostiainen 15:00:45 Regrets+ Markus_Tavenrath 15:00:45 RRSAgent, draft minutes 15:00:46 I have made the request to generate https://www.w3.org/2026/10/08-webmachinelearning-minutes.html anssik 15:01:14 Present+ Mike_Wyrzykowski 15:01:29 Present+ Ningxin_Hu 15:01:38 Present+ Reilly_Grant 15:01:42 ningxin has joined #webmachinelearning 15:01:59 mklimenko has joined #webmachinelearning 15:02:19 Present+ Sushanth_Rajasankar 15:02:24 Present+ Dwayne_Robinson 15:03:09 lgombos has joined #webmachinelearning 15:03:30 sushraja-msft has joined #webmachinelearning 15:03:47 RafaelCintron has joined #webmachinelearning 15:04:05 Present+ Mikhail_Klimenko 15:04:13 Ehsan has joined #webmachinelearning 15:04:18 Present+ Christian_Liebel 15:04:25 Present+ Ehsan_Toreini 15:04:47 ... we have a new Zoom meeting setup and due to that you should see a prompt about transcription when you join 15:04:54 Has the meeting URL changed? 15:04:58 ... 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 15:05:49 https://w3c.zoom.us/j/82153370456?pwd=uyFek0b30VKsrYlPn4NIJIJ8segxwC.1 15:06:34 s|https://w3c.zoom.us/j/82153370456?pwd=uyFek0b30VKsrYlPn4NIJIJ8segxwC.1|| 15:06:55 Anssi: welcome to Dave Raggett from W3C joining the WebML WG 15:07:47 Topic: Web Neural Network API 15:07:58 gb, this is webmachinelearning/webnn 15:07:59 anssik, OK. 15:08:12 Anssi: F2F agenda has been published: 15:08:16 -> WebML WG/CG F2F Agenda - TPAC 2026 (Dublin, Ireland) https://github.com/webmachinelearning/meetings/issues/39 15:08:17 https://github.com/webmachinelearning/meetings/issues/39 -> Issue 39 WebML WG/CG F2F Agenda - TPAC 2026 (Dublin, Ireland) (by anssiko) [Agenda+] 15:08:24 Anssi: our plan is as follows 15:08:29 ... GMT morning session: 15:09:11 ... - current baseline, major milestones since Kobe 15:09:17 ... - proposed breaking changes 15:09:53 ... - device selection & effective compute policy 15:10:23 ... this dedicated session is in response to the customer demand for a mechanism to inform client and cloud execution decisions 15:10:58 ... after the first break, LLM enablement 15:11:21 ... - dynamic shapes next steps with implementation experience 15:11:37 ... - subgraphs and operator set 15:12:04 ... - low precision data types and quantization 15:12:29 ... after lunch implementation experience and ecosystem 15:12:51 ... - demos and discussion on implementation experience 15:13:08 ... - customer stories that connect our work to real-world use cases 15:13:39 ... - prototyping streaming constants and caching compiled graphs 15:14:09 ... as for the WebGPU interop v2, Bryan is on vacation during F2F, so consider deferring this, unless we have WebGPU experts available 15:14:25 ... MikeW, anyone from WebGPU land TPAC? 15:14:49 MikeW: I might participate remotely 15:15:08 Rafael: I will be at WebGPU F2F next week after TPAC, remotely at TPAC 15:15:29 ... after afternoon break 15:15:43 ... - security hardening 15:16:22 ... - numerics & WPT 15:16:49 ... - any horizontal review topics and wrap up 15:16:58 ... close with a dinner 15:17:00 q? 15:17:53 Christian: TAG has not talked about WebNN recently so I don't think we need a TAG-WG discussion at TPAC 15:18:20 Anssi: I acknowledge that is a lot to cover in one day 15:18:57 ... I will consider your feedback and will prune the agenda accordingly 15:20:05 Anssi: are you involved with privacy reviews? 15:20:16 Ehsan: nothing on top of my head, will check 15:21:40 q? 15:22:28 Subtopic: Low-precision floating point data types 15:22:41 Anssi: issue #930, explainer PR #938 and spec PR #956 15:22:42 https://github.com/webmachinelearning/webnn/pull/938 -> PR 938 Add low-precision floating point data types explainer (by mklimenko-nv) [Agenda+] 15:22:42 https://github.com/webmachinelearning/webnn/issues/930 -> Issue 930 RFE: Add support for more floating point low-precision ML data types (`bfloat16`, `fp8`, `nvfp4`) (by mtavenrath) [opset] [feature request] [Agenda+] 15:22:43 https://github.com/webmachinelearning/webnn/pull/956 -> PR 956 Add `bfloat16` and `float8e4m3` to the spec (by mklimenko-nv) [Agenda+] 15:22:56 ... explainer PR approved by Dwayne, spec PR review submitted by Dwayne 15:23:10 ... I invite the group to review the PRs and thank Mikhail for all the contributions 15:23:35 ... I have put low-precision data types next steps to the F2F agenda, so overflow from these PRs will get some F2F time 15:24:13 Mikhail: great feedback from Dwayne, hopefully we can discuss this at F2F 15:24:27 Present+ Laszlo_Gombos 15:25:16 Subtopic: Conv2d groups 15:25:21 Anssi: issue #959 15:25:22 https://github.com/webmachinelearning/webnn/issues/959 -> Issue 959 conv2d: infer groups from input and filter shapes when not specified (by ibelem) [Agenda+] 15:25:35 ... Belem reports MLConv2dOptions.groups IDL defaults to 1 15:26:07 ... this means grouped convolutions only validate if groups is passed explicitly 15:26:22 ... in these implementations, group count is implied via other information 15:26:35 ... the problem is implementations that map TFLite ops 1:1 to WebNN fail validation for grouped convolutions 15:26:58 ... 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 15:27:45 Dwayne: not opposed to the change, but is this necessary? 15:27:57 ... if TFLite delegate calls into WebNN it could do this computation before reaching WebNN 15:28:06 q+ 15:28:08 ack ningxin 15:29:06 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 strict validation on this 15:29:07 q? 15:29:36 Dwayne: we need to understand this issue a bit more 15:30:31 q? 15:30:55 Anssi: we'll come back to this at F2F if new information emerges 15:31:11 Subtopic: batchNormalization + gelu error tolerance 15:31:16 Anssi: issue #950 15:31:16 https://github.com/webmachinelearning/webnn/issues/950 -> Issue 950 Discrepancy in actual and expected output for subgraph test : batchNormalization options.axis=0 + gelu (by hthadicherla) [Agenda+] [web-platform-tests] 15:31:30 ... implementation feedback from TensorRT-RTX EP with ORT backend 15:31:45 ... given a graph with batchNormalization options.axis=0 + gelu 15:31:51 ... the strictness checks for batchNormalization + gelu fail error tolerance 15:32:03 ... also other tested backends fail this strictness check, ORT CPU, TFLite, LiteRT GPU 15:32:15 q? 15:33:04 Dwayne: this would need to 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 15:33:06 q? 15:33:56 Anssi: spec would remain as is and this only suggests WPT changes? 15:34:00 Dwayne: correct 15:34:00 q? 15:35:01 Subtopic: Full-sequence output validation 15:35:11 Anssi: issue #949 and PR #955 15:35:11 https://github.com/webmachinelearning/webnn/issues/949 -> Issue 949 Validate LSTM/GRU full-sequence output regardless of the `returnSequence` flag (by shiyi9801) [security-tracker] 15:35:11 https://github.com/webmachinelearning/webnn/pull/955 -> PR 955 Enforce validation when creating an MLOperandDescriptor (by shiyi9801) 15:35:19 ... we have a resolution for this issue: 15:35:24 -> https://www.w3.org/2026/09/10-webmachinelearning-minutes.html#7460 15:35:44 Anssi: PR submitted by Shiyi reviewed by Ningxin, any additional feedback or comments? 15:35:50 ... we can merge this PR once all feedback is addressed. 15:37:18 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 do this right for all affected ops 15:38:21 Anssi: two LGTMs, ready to land 15:38:46 Subtopic: Accuracy for reductions 15:38:52 Anssi: issue #947 and PR #957 15:38:53 https://github.com/webmachinelearning/webnn/issues/947 -> Issue 947 reduceLogSumExp reference decomposition contradicts conformance test; accuracy requirements missing (by mtavenrath) [web-platform-tests] 15:38:53 https://github.com/webmachinelearning/webnn/pull/957 -> PR 957 Add stable decomposition clarification (by mklimenko-nv) 15:38:56 ... similarly, we have a resolution for this: 15:39:00 -> https://www.w3.org/2026/09/10-webmachinelearning-minutes.html#214b 15:39:36 Anssi: and spec PR is reviewed by Dwayne with minor comments, Ningxin's review requested 15:39:38 I'll take a look 15:39:41 ... we can merge also this PR once all feedback is addressed 15:40:21 Dwayne: just wanted to add a minor comment why we want to have this optimization 15:41:03 Mikhail: fair comment from Dwayne, apologies for missing this, will fix this 15:41:27 Subtopic: Next meeting is TPAC F2F 15:42:27 Anssi: due to close proximity of the TPAC F2F meeting, please consider this an opportunity to discuss any pending issues in person 15:42:35 ... our bi-weekly meeting on 22 October 2026 canceled 15:43:42 ... looking forward to seeing many of you in Dublin later this month 15:44:12 ... safe travels, and please remember to review the F2F agenda that is shaping up, and submit your comments: 15:44:17 -> WebML WG/CG F2F Agenda - TPAC 2026 (Dublin, Ireland) https://github.com/webmachinelearning/meetings/issues/39 15:44:18 https://github.com/webmachinelearning/meetings/issues/39 -> Issue 39 WebML WG/CG F2F Agenda - TPAC 2026 (Dublin, Ireland) (by anssiko) [Agenda+] 15:44:36 q? 15:45:26 RRSAgent, draft minutes 15:45:27 I have made the request to generate https://www.w3.org/2026/10/08-webmachinelearning-minutes.html anssik 15:48:27 s/WebGPU land TPAC/WebGPU-land at TPAC" 15:48:37 s/TPAC"/TPAC 15:49:57 s/are you involved with privacy reviews?/Ehsan, are you involved with privacy reviews, any privacy topics on your list? 15:52:22 s/does not strict/does not do strict 15:53:24 s/to relaxed/to be relaxed 15:54:02 s/do this right/this done right 15:57:06 s/Anssi: F2F agenda has been published:/Subtopic: TPAC F2F plan for WebNN/ 15:57:18 RRSAgent, draft minutes 15:57:20 I have made the request to generate https://www.w3.org/2026/10/08-webmachinelearning-minutes.html anssik 16:07:52 s/canceled/is canceled 16:08:05 s/consider this/consider TPAC 16:08:07 RRSAgent, draft minutes 16:08:08 I have made the request to generate https://www.w3.org/2026/10/08-webmachinelearning-minutes.html anssik 18:00:29 Zakim has left #webmachinelearning