14:58:55 RRSAgent has joined #webmachinelearning 14:59:00 logging to https://www.w3.org/2026/09/24-webmachinelearning-irc 14:59:00 RRSAgent, make logs Public 14:59:01 please title this meeting ("meeting: ..."), anssik 14:59:02 Meeting: WebML WG Teleconference – 24 September 2026 14:59:06 Chair: Anssi 14:59:11 Agenda: https://github.com/webmachinelearning/meetings/blob/main/telcons/2026-09-24-wg-agenda.md 14:59:19 Scribe: Anssi 14:59:20 scribeNick: anssik 14:59:29 Present+ Anssi_Kostiainen 14:59:35 RRSAgent, draft minutes 14:59:36 I have made the request to generate https://www.w3.org/2026/09/24-webmachinelearning-minutes.html anssik 15:00:23 Present+ Dwayne_Robinson 15:00:59 Present+ Bryan_Bernhart 15:01:22 Present+ Ningxin_Hu 15:01:54 Present+ Reilly_Grant 15:02:11 DwayneR has joined #webmachinelearning 15:03:05 RRSAgent, draft minutes 15:03:06 I have made the request to generate https://www.w3.org/2026/09/24-webmachinelearning-minutes.html anssik 15:03:58 Present+ Markus_Tavenrath 15:05:02 RRSAgent, draft minutes 15:05:04 I have made the request to generate https://www.w3.org/2026/09/24-webmachinelearning-minutes.html anssik 15:05:14 Topic: Web Neural Network API 15:05:19 gb, this is webmachinelearning/webnn 15:05:19 anssik, OK. 15:05:24 Subtopic: WebNN-WebGPU interop v2 features 15:05:35 Anssi: initial WebNN-WebGPU interop feature was merged in PR #942 15:05:36 https://github.com/webmachinelearning/webnn/pull/942 -> MERGED PR 942 Add initial WebNN-WebGPU Interop specification (by bbernhar) 15:05:41 ... as discussed last time, we'll enhance this base with v2 features 15:05:54 ... first v2 proposal is zero-copy hints, issue #941 15:05:55 https://github.com/webmachinelearning/webnn/issues/941 -> Issue 941 WebNN/WebGPU zero-copy device selection (by bbernhar) [webgpu interop] [Agenda+] 15:06:22 Anssi: we had emerging consensus last time to use the interopDevice option as in createContext({ interopDevice: gpuDevice }) 15:06:27 ... I see support from Reilly, Bryan and MikeW 15:06:43 ... any concerns in resolving to advance with this approach? 15:06:44 q? 15:07:53 RafaelCintron has joined #webmachinelearning 15:08:05 Bryan: we wanted to allow use the same dedicated GPU for interop, an improvement from the prior state 15:08:09 ... Reilly proposal is good 15:08:17 Ehsan has joined #webmachinelearning 15:08:36 Present+ Ehsan_Toreini 15:08:53 Anssi: any implementation plans? 15:09:14 Bryan: there's a plan, multi-adapter not fully fleshed out in WebGPU-land, so we have a dependency on that 15:09:45 MarkusT: would WebGPU [inaudible] device selection? 15:10:45 MarkusT: I will review the issue and come back to it next week 15:11:21 Anssi: MarkusT will share feedback by our next meeting in the issue 15:11:22 q? 15:11:57 Anssi: second v2 proposal, MLTensor export custom usages & flags issue #954 15:11:58 https://github.com/webmachinelearning/webnn/issues/954 -> Issue 954 MLTensor export to GPUBuffer with custom usages? (by mwyrzykowski) [webgpu interop] [Agenda+] 15:12:20 ningxin has joined #webmachinelearning 15:12:39 Anssi: Dwayne's flag usage was recast on the Mike's proposed approach 15:13:06 Dwayne: I think that Bryan's latest comment makes sense, just using GPUBufferUsageFlags 15:13:07 q? 15:14:01 Bryan: do we want WebNN usage to match, two different approach, strongly typed and in WebGPU a different approach 15:14:24 Dwayne: question, any WebGPU flags that would not make sense to inherit? 15:15:00 Reilly: I through are there any flags, because need to allocate to NPU, GPU flags that could't be used to create a GPU buffer out of that buffer 15:15:11 Anssi: notes Dwayne's +1 15:15:39 Reilly: I'm not familiar enough with WebGPU to answer that, an area to look into as we prototype this feature 15:16:28 Bryan: if you create mappable GPU buffer, don't want to always apply that, GPU bandwidth consideration, prefer minimal set of flags 15:16:45 ... explicit usages are useful and avoid implicit behavior that requires additional copies 15:17:21 Reilly: exported tensor has less of those than GPU buffer, you already have read tensor and send data to GPU 15:18:07 Bryan: we found mappable buffers useful without interop also, a questions if a developer want to use mappable memory, is interop path mandatory? 15:18:10 q? 15:18:24 Present+ Rafael_Cintron 15:18:38 Reilly: why you want those buffers 15:19:03 s/those buffers/mappable buffers 15:19:18 Bryan: write and read buffer semantically asks to pass a copy 15:20:03 ... workloads that do a lot of initialization can spare that, rather than create a JS buffer or array buffer, also UMA optimization at your disposal 15:20:45 ... readback has no no dependencies, write implicitly says I want a copy, APIs not great at mapped rights, WebGPU has mapAsync and map APIs separately for this purpose 15:21:06 ... do we want to support mapping, GPU usages can be applied when they make sense, maybe in non-interop scenarios too 15:22:03 Reilly: should we have a map on map capability in MLTensor API, obviously we can prototype the benefits to see if it is a thing to add to WebNN directly without going through WebGPU 15:22:29 ... there are systems that jump through WebGPU for a buffer that does not exists 15:24:10 Subtopic: Accumulation precision 15:24:15 Anssi: issue #948 15:24:16 https://github.com/webmachinelearning/webnn/issues/948 -> Issue 948 Accumulation precision is unspecified for matmul/gemm and reduction operations (by mtavenrath) [Agenda+] 15:24:22 ... MarkusT wanted to follow up on this issue we discussed last time on our Aug meeting 15:24:26 -> https://www.w3.org/2026/08/27-webmachinelearning-minutes.html#c2ec 15:24:32 Anssi: two solutions were proposed: 15:24:38 ... A) explicit upcast 15:24:44 ... B) explicit accumulator type per operation 15:24:51 ... Ningxin had a question if we emulate option B 15:24:55 ... MarkusT responded that this should be minimum accumulator types, no need to emulate 15:25:33 MarkusT: upcast is what ONNX does by default, I don't like A because it requires transmitting the graph to WebNN 15:26:07 ... option B, give default and specify precision per op 15:26:51 ... I initially wrote this in a way we can support certain number of bits for exponent and mantissa 15:27:21 ... for example, if we expose precision on NVIDIA hardware, get 32 exponent and only a few bits for mantissa 15:28:41 q+ 15:28:45 MarkusT: approach B in #948 reflects my current recommendation on how to address this issue, comments? 15:28:54 q? 15:29:10 ack reillyg 15:29:38 Reilly: I see a parallel between this and quantization 15:30:08 ... for WebNN, we chose to follow the model of frameworks like ONNX RT with explicitly quantized ops, and dequantized after 15:30:27 ... cf. LiteRT with this built-in 15:31:22 ... if you combine this with accumulator type being explicit, you could have quantization attached to tensors 15:31:40 If B, then I'd rather have any accumulator MLDataType exposed publicly as enums than passing raw bit counts for mantissa. 15:31:53 Reilly: most explicit way to express for each value 15:32:34 ... agree with Dwayne's comment above, all data types should be well defined 15:32:52 ... passing explicit mantissa makes an assumption it is floating point that no each operator wants 15:33:11 MarkusT: I read this you are saying option B and enhancing quantization? 15:33:17 ... and backends do upcast as required? 15:33:31 ... Reilly, do you agree with option B? 15:33:47 Reilly: I agree with option B but is raises questions around how we represent quantization 15:34:25 ... given the quantization design, I wonder whether we lose expressive power by using that design, and what we lose to just use upcast for accumulator precision 15:35:07 ... question to MarkusT, in TensorRT, I assume there's some compiler optimization pass that happens where the cast is moved inside the op? 15:35:32 ... accumulator only affected, not the precision of the in-memory tensors? 15:35:47 MarkusT: correct 15:36:51 ... I think operator casting is not so nice design 15:38:15 Reilly: I appreciate the specificity of approach B, and generalizability of specifying just precision bits 15:38:36 ... only question is, if we choose approach B, should consider the same design for quantized tensors 15:39:16 Dwayne: option B would cascade through the whole API 15:39:36 ... for quantization, if we bundle that to every MLTensor can become complicated 15:40:00 Reilly: quantization information into the tensor makes data types checks more complicated 15:40:51 ... we can define it in WebNN, quantization tensor passed to operator not supporting operation on such tensor, every operator can operate on quantized data 15:41:14 ... quantizations only make sense on float value, so implicitly those won't take quantization values 15:41:31 ... those ops don't take floats, 16, 32, others 15:42:24 ... this implementation complexity is in implementation already, due to fusing, e,.g. in LiteRT we have quantized tensors, and similarly ONNX has code to do the same in reverse when we see quantized operators 15:42:39 ... adds complexity to API but minimal implementation complexity increase 15:43:33 q+ 15:43:40 ... e.g. in matMul, specify both accumulator precision and preferred output quantization, those end up on the operator, not sure if that'd be a bad thing, makes things explicit 15:43:41 q? 15:44:14 MarkusT: I don't think we need to add this information every op, e.g. min/max, only to those that reduce values in the end 15:44:30 Reilly: memory movement ops would be defined passing through any data type they get 15:45:04 MarkusT: gemm and convolutions most affected, other option to state this globally 15:45:33 ... in TensorRT, fp32 gemms kills some bits from internal floats to get faster maths 15:45:45 s/kills/drop 15:46:03 q? 15:46:29 ack ningxin 15:47:01 Ningxin: a question regarding discoverability, two parts, i) implementability, the proposal is to add accumulator type to each op that support this 15:47:29 ... per the table this is not per op granularity, rather a global type 15:47:53 croissant` has joined #webmachinelearning 15:47:59 ... ii) developer's perspective, developer checks accummulator type and it is not supported, what to do, use cast instead? 15:48:33 MarkusT: for (ii) if not supported and developer uses cast, can we do the same thing that the Chromium backend does? 15:48:57 ... depending on the backend, either upcast or pass the accumulator type to the underlying implementation 15:49:17 MarkusT: developer can specify the desired accumulator type, implementation detail how this is applied 15:49:42 ... we don't define accumulator type at all, if the developer wants to go higher, need to upcast 15:50:07 ... don't force the developer to use upcast, but allow use an appropriate accumulator type 15:50:24 Ningxin: question about discoverability 15:51:10 Ningxin: if the backend is browser can always do casting 15:51:47 MarkusT: a different explicit way, instead of upcast to float32 15:51:48 q? 15:52:43 MarkusT: if no objection, can do a detailed explainer for approach B to address feedback 15:55:00 +1 15:55:02 RESOLUTION: Create an explainer for accumulator precision based on approach B that addresses feedback from the group. (issue #948) 15:55:40 are we ready to merge the dynamic shape PR? 15:55:49 Subtopic: Dynamic shapes 15:55:54 Anssi: issue #883 15:55:54 https://github.com/webmachinelearning/webnn/issues/883 -> Issue 883 Support flexible input sizes (by huningxin) [feature request] [operator specific] 15:56:06 Anssi: this meeting was the last call for comments 15:56:16 ... thank you Dwayne for your final review and suggestions for humanization of the text 15:56:28 ... it really shows the value of a human review 15:56:49 ... my expectation is all contributors will conduct a human review before submitting to catch word salad issues but for non-English speakers this might be particularly challenging, so appreciate native speakers' help 15:57:00 ... after Dwayne's final review comments are addressed, the editors can merge this PR 15:57:26 ... thank you Miao for swift handling of the review comments 15:57:32 q? 15:58:10 +1 15:58:21 RESOLUTION: The group will merge the Dynamic Shape Explainer PR after addressing final review comments and proceeds to implementation experience and spec PR phase next. (issue #883) 15:58:26 RRSAgent, draft minutes 15:58:27 I have made the request to generate https://www.w3.org/2026/09/24-webmachinelearning-minutes.html anssik 15:59:26 -> TPAC WIP agenda https://github.com/webmachinelearning/meetings/issues/39 15:59:27 https://github.com/webmachinelearning/meetings/issues/39 -> Issue 39 WebML WG/CG F2F Agenda - TPAC 2026 (Dublin, Ireland) (by anssiko) 16:00:14 RRSAgent, draft minutes 16:00:15 I have made the request to generate https://www.w3.org/2026/09/24-webmachinelearning-minutes.html anssik 16:08:27 s/Reilly proposal/Reilly's proposal 16:09:31 s/two different approach/two different approaches 16:10:25 s/through/thought 16:11:36 s/no no/no 16:12:47 s/like A/like approach A 16:13:11 s/give default and specify/gives default and specifies 16:14:25 s/is raises/it raises 16:15:24 s/can become/this can become 16:15:50 s/into the/in the 16:19:24 RRSAgent, draft minutes 16:19:25 I have made the request to generate https://www.w3.org/2026/09/24-webmachinelearning-minutes.html anssik 18:00:52 Zakim has left #webmachinelearning