14:58:09 <RRSAgent> RRSAgent has joined #webmachinelearning
14:58:13 <RRSAgent> logging to https://www.w3.org/2026/08/27-webmachinelearning-irc
14:58:13 <Zakim> RRSAgent, make logs Public
14:58:14 <Zakim> please title this meeting ("meeting: ..."), anssik
14:58:16 <anssik> Meeting: WebML WG Teleconference – 27 August 2026
14:58:22 <anssik> Chair: Anssi
14:58:26 <anssik> Agenda: https://github.com/webmachinelearning/meetings/blob/main/telcons/2026-08-27-wg-agenda.md
14:58:31 <anssik> Scribe: Anssi
14:58:36 <anssik> scribeNick: anssik
14:58:49 <anssik> Present+ Anssi_Kostiainen
14:58:55 <anssik> RRSAgent, draft minutes
14:58:57 <RRSAgent> I have made the request to generate https://www.w3.org/2026/08/27-webmachinelearning-minutes.html anssik
14:59:42 <DwayneR> DwayneR has joined #webmachinelearning
15:00:26 <mklimenko> mklimenko has joined #webmachinelearning
15:01:17 <anolan> anolan has joined #webmachinelearning
15:01:45 <anssik> Present+ Ruoxi_Ran
15:01:50 <anssik> Present+ Bryan_Bernhart
15:01:59 <anssik> Present+ Dwayne_Robinson
15:02:09 <anssik> Present+ Mikhail_Klimenko
15:02:24 <anssik> Present+ Thomas_Steiner
15:02:55 <anssik> RRSAgent, draft minutes
15:02:56 <RRSAgent> I have made the request to generate https://www.w3.org/2026/08/27-webmachinelearning-minutes.html anssik
15:03:11 <mtavenrath> mtavenrath has joined #webmachinelearning
15:03:18 <anssik> Present+ Christian_Liebel
15:03:28 <anssik> Present+ Rafael_Cintron
15:03:36 <anssik> Present+ Reilly_Grant
15:03:54 <anssik> Present+ Ningxin_Hu
15:04:02 <alispivak> alispivak has joined #webmachinelearning
15:04:07 <ningxin> ningxin has joined #webmachinelearning
15:04:07 <anssik> Present+ Ali_Spivak
15:04:13 <anssik> ... welcome to the new participants joining us since our last meeting:
15:04:16 <RafaelCintron> RafaelCintron has joined #webmachinelearning
15:04:18 <anssik> ... Jay Mok from PayPal
15:04:22 <anssik> ... Mikhail Klimenko from NVIDIA
15:04:30 <anssik> ... Thomas Steiner and Rezvan Mahdavi Hezaveh from Google
15:04:35 <anssik> ... welcome all!
15:04:43 <anssik> Topic: New incubation: Cross-Origin Storage
15:05:06 <anssik> gb, this is wicg/cross-origin-storage
15:05:06 <gb> anssik, OK.
15:05:27 <anssik> -> Cross-Origin Storage spec https://wicg.github.io/cross-origin-storage/
15:06:14 <anssik> Thomas: content-addressable browser cache, shares files across origins by hash
15:06:57 <anssik> ... currently, caches are partitioned in all browsers
15:07:45 <anssik> ... huge potential for sharing commonly used resources, e.g. AI models, Wasm files, Web fonts, Game engines, libraries
15:07:54 <DwayneR> DwayneR has joined #webmachinelearning
15:09:05 <anssik> ... the proposal is called Cross-Origin Storage (COS), Blob-based, resource hashes, extends FileSystemFileHandle, visibility limited to certain origins per allow-list
15:09:34 <anssik> [proposed API code sample, imperative]
15:11:02 <anssik> s/sample/sample for reading
15:11:13 <anssik> [proposed API code sample for writing, imperative]
15:11:53 <anssik> Thomas: declarative approach using <link> and <script> extensions
15:12:07 <anssik> ... declarative approach for CSS also
15:12:42 <anssik> ... integration with fetch() with extensions
15:13:05 <anssik> -> https://goo.gle/cos-extension
15:13:24 <anssik> Thomas: demo available for testing
15:14:22 <anssik> ... Chromium build available with COS API implemented including DevTools support:
15:14:40 <anssik> [link to build]
15:15:21 <anssik> Thomas: HuggingFace, MLC AI WebLLM, wllama, LiteRT-ML JS, Flutter, Emscripten etc. as initial customers
15:15:52 <anssik> ... Chrome team would like to make this a proper API
15:16:13 <anssik> -> https://goo.gle/cos-slides
15:16:36 <Roy> Roy has joined #webmachinelearning
15:16:48 <anssik> s|[link to build]|https://goo.gle/cos-extension
15:17:16 <anssik> [drum roll, COS demo unveiling]
15:17:40 <anssik> Thomas: patched webnn-samples repo for this COS WebNN demo
15:18:23 <anssik> ... cache miss/hit shown in the Chrome UI
15:19:23 <anssik> ... this is true cross-origin shared resource, the AI model used for the style transfer sample
15:19:51 <anssik> -> https://web-ai-community.github.io/webnn-samples/style_transfer/
15:19:53 <anssik> -> https://tomayac.github.io/webnn-samples/style_transfer/
15:19:58 <anssik> -> https://github.com/tomayac/awesome-cross-origin-storage
15:20:47 <anssik> Thomas: DevTools support allows to inspect cached files, AI models in this case
15:23:10 <anssik> q?
15:23:11 <mtavenrath> q+
15:23:13 <RafaelCintron> q+
15:23:45 <anssik> Thomas: this proposal is general purpose so it will likely go to more general purpose group for standardization
15:23:59 <anssik> ... WebNN is one important customer for this API
15:24:06 <anssik> ack mtavenrath
15:24:31 <anssik> MarkusT: very cool, does the file have to match the hash, do I have to compute hash first?
15:25:06 <ningxin> ningxin has joined #webmachinelearning
15:25:10 <anssik> Thomas: HF publishes hashes for the models they host, there's a public hash list on their site
15:25:23 <Roy_Ruoxi> Roy_Ruoxi has joined #webmachinelearning
15:26:40 <anssik> MarkusT: if you store, from where does the hash come from?
15:27:08 <ningxin> ningxin has joined #webmachinelearning
15:27:17 <anssik> Thomas: hash comes from Hugging Face Hub
15:28:14 <anssik> MarkusT: would be nice if the UI would provide a mechanism for the user to clean up cache as appropriate
15:29:26 <anssik> Thomas: in Chrome you can remove the site's data, does not differentiate on a file basis
15:30:37 <anssik> ... a regular user does not need to case, the implementation handles that for the user
15:30:38 <anssik> q?
15:30:40 <anssik> ack RafaelCintron
15:31:25 <anssik> Rafael: first question was, request blob and you'll get a hash, how does the implementation know what to look for given the hash?
15:32:06 <anssik> Thomas: there needs to be a URL that is mapped to the hash initially
15:32:19 <anssik> ... URL is to deliver the resource, then forgotten
15:32:35 <anssik> Rafael: contents of the file does not matter?
15:32:44 <anssik> Thomas: if the file content changes, the hash changes
15:32:48 <anssik> q?
15:33:05 <anssik> Rafael: internally the browser keeps a checksum?
15:33:44 <anssik> Thomas: right, this ensures integrity
15:34:29 <anssik> Rafael: about cross-origin, when you list URLs, can I use wildcards (*)?
15:34:43 <Ehsan> Ehsan has joined #webmachinelearning
15:35:02 <anssik> Thomas: our design philosophy is "people will do stupid things"
15:35:37 <anssik> ... core idea is you need to use public hashes
15:35:47 <anssik> q?
15:36:28 <anssik> -> https://github.com/WICG/cross-origin-storage/blob/main/public-hash-list/phl-explainer.md
15:37:38 <anssik> -> Prototype browser implementations with native support https://github.com/tomayac/awesome-cross-origin-storage#prototype-browser-implementations
15:37:42 <anssik> -> Extension for all major browsers https://github.com/web-ai-community/cross-origin-storage-extension#available-on
15:38:03 <anssik> q?
15:38:55 <anssik> Topic: Web Neural Network API
15:38:59 <anssik> gb, this is webmachinelearning/webnn
15:38:59 <gb> anssik, OK.
15:39:03 <anssik> Subtopic: WebGPU interop
15:39:08 <anssik> Anssi: PR #945
15:39:08 <gb> https://github.com/webmachinelearning/webnn/pull/945 -> PR 945 Add dynamic shape explainer (by miaobin)
15:39:27 <anssik> Anssi: in our June meeting we resolved to port over the necessary features to the WebNN API spec from the MLTensor explainer
15:39:31 <anssik> -> https://www.w3.org/2026/06/18-webmachinelearning-minutes.html#2e7b
15:39:55 <anssik> Anssi: Bryan has shared with the group PR #942 for the proposed spec updates, ready for review
15:39:56 <gb> https://github.com/webmachinelearning/webnn/pull/942 -> PR 942 Add initial WebNN-WebGPU Interop specification (by bbernhar) [Agenda+]
15:40:03 <anssik> s/PR #945/PR #942
15:40:08 <anssik> Anssi: the PR adds the following features to the WebNN API spec:
15:40:16 <anssik> -> createExportableTensor() https://pr-preview.s3.amazonaws.com/bbernhar/webnn/pull/942.html#api-mlcontext-createexportabletensor
15:40:20 <anssik> -> exportToGPU() https://pr-preview.s3.amazonaws.com/bbernhar/webnn/pull/942.html#api-mlcontext-exporttogpu
15:40:45 <anssik> Anssi: createExportableTensor() creates an MLTensor associated with this MLContext which may be exported to the given GPUDevice using exportToGPU()"
15:40:49 <anssik> Anssi: exportToGPU() exports an exportable MLTensor to its associated GPUDevice as a GPUBuffer, and rents it to that GPUDevice until the returned GPUBuffer is destroyed
15:40:53 <anssik> ... these spec changes follow the explainer with minor naming adjustment (createTensor -> createExportableTensor):
15:40:56 <anssik> -> https://github.com/webmachinelearning/webnn/blob/main/mltensor-explainer.md#webgpu-interop
15:41:04 <anssik> Anssi: Bryan, please walk us through the PR and any open questions you'd like the group to discuss
15:41:14 <anssik> ... after that, we'll discuss the feedback shared via GH from MikeW, Ningxin, and any other comments from the group
15:41:22 <anssik> q?
15:41:57 <anssik> Bryan: this is an initial spec PR we've used in our implementation, there are likely some edge cases, open questions on multi-adapter scenarios
15:42:11 <anssik> ... in terms of basic handoff API, zero-copy enabling, it is more or less there
15:42:19 <anssik> ... to support more formats is future work
15:42:44 <anssik> ... this is the first draft for basic interop, the spec PR is self-explanatory, uses strictly WebGPU types
15:43:01 <anssik> ... interop tensor looks a lot like normal tensor and you can use it in WebGPU directly
15:43:16 <anssik> ... not a lot of additional API exposed to web developers
15:43:26 <RafaelCintron> q+
15:43:26 <anssik> q?
15:43:41 <anssik> Bryan: feedback on the PR directly welcome
15:43:45 <anssik> ack RafaelCintron
15:44:19 <anssik> Rafael: I had a couple of questions about the PR, sometimes middleware developers don't fully know how their middleware is used, e.g. ThreeJS in WebGPU world
15:44:33 <anssik> ... make all tensors in their world exportable, can have performance implications
15:44:53 <anssik> ... should be give advice to web developers about this
15:45:16 <anssik> ... what if the chosen adapter does not match?
15:45:25 <anssik> q?
15:46:15 <ningxin> q+
15:46:17 <anssik> Bryan: not in developers' hands, use interop when you need it
15:46:29 <anssik> ... we can warn developers of performance cliffs
15:47:02 <anssik> Rafael: on the surface it'd be cool if you could say high power, and if. you could pass GPU device and have it work, have the browser copy for you
15:47:24 <anssik> ... performance cliff possible if browser picks a wrong adapter that needs copies
15:48:23 <anssik> Bryan: high performance is overloaded, an open issue to spec for the source and destination devices to match
15:48:48 <anssik> ... a proposal how device selection is influenced, orthogonal, current spec v1 assumes this just works
15:48:48 <anssik> q?
15:48:52 <anssik> ack ningxin
15:49:27 <anssik> Ningxin: there are some webgpu-interop open issues, could the spec PR embed those issues?
15:49:59 <anssik> ... about narrow scoped usage, when exporting a tensor, you can still read and write to it, should we restrict that?
15:50:35 <anssik> Bryan: for read/write, we could, but would require developers to go to WebGPU for that
15:51:12 <anssik> Ningxin: are there performance implications to make a tensor read/write, use a different memory region?
15:51:48 <anssik> Bryan: if you specify a GPU device and we make it not readable and/or writable, there are performance implications
15:52:00 <anssik> ... creating MLContext, how it can influence the destination buffer
15:52:01 <anssik> q?
15:53:16 <anssik> Rafael: I will review the PR
15:54:19 <anssik> Reilly: I'm happy to see this spec'd, agree with earlier discussion on performance cliff mitigation, we can probably make rules that address this issue
15:54:20 <anssik> q?
15:55:03 <anssik> Subtopic: Accumulation precision
15:55:10 <anssik> Anssi: issue #948
15:55:10 <gb> https://github.com/webmachinelearning/webnn/issues/948 -> Issue 948 Accumulation precision is unspecified for matmul/gemm and reduction operations (by mtavenrath) [Agenda+]
15:55:34 <anssik> Anssi: MarkusT noted the spec does not say anything about what precision intermediate accumulations should use
15:56:11 <anssik> MarkusT: matmul and gemm the most significantly affected, but also other ops that sum values internally are affected e.g. reduceSum, conv2d, gru, lstm etc.
15:56:23 <anssik> ... what is the minimum precision the group would support?
15:56:40 <anssik> ... is it exactly same as that of inputs?
15:56:55 <anssik> ... or should we add an extra attribute that specified the accumulator type
15:57:10 <anssik> MarkusT: A) explicit upcast
15:57:16 <anssik> ... cast inputs to float32, perform the operation, and cast back
15:57:22 <anssik> MarkusT: B) explicit accumulator type per operation
15:57:26 <anssik> ... add an optional accumulatorType field to operations that have internal accumulation
15:57:44 <anssik> MarkusT: which approach is preferred?
15:57:45 <anssik> q?
15:58:52 <anssik> Dwayne: I like what I see, will provide more input
15:59:25 <ningxin> q+
15:59:28 <anssik> ack ningxin
15:59:49 <anssik> Ningxin: I think this is important issue, we see this in e.g. diffusion models
16:00:01 <anssik> ... if you spec proposal B, is there composition or emulation path available?
16:00:16 <anssik> MarkusT: we don't need any emulation, this should be minimum accumulation type
16:01:05 <anssik> Ningxin: makes sense, if e.g. polyfill sees this option, what does it do, should it go through the cast pass to emulate this?
16:01:14 <anssik> MarkusT: not sure about that case
16:01:24 <anssik> Ningxin: I will comment on the issue
16:01:51 <anssik> MarkusT: considering WebGPU, could we have WebGPU implementation working on an exact accumulator type
16:02:30 <anssik> ... how to ensure people can test the model runs in such a scenario
16:02:47 <anssik> ... we need to define the WebGPU implementation uses minimum accumulator type anywhere
16:02:48 <anssik> q?
16:04:11 <anssik> RRSAgent, draft minutes
16:04:12 <RRSAgent> I have made the request to generate https://www.w3.org/2026/08/27-webmachinelearning-minutes.html anssik
16:04:57 <anssik> Present+ Andrew_Nolan
16:05:19 <anssik> Present+ Markus_Tavenrath
16:06:23 <anssik> Present+ Rezvan_Mahdavi_Hezaveh
16:06:49 <anssik> Present+ Ehsan_Toreini
16:07:10 <anssik> RRSAgent, draft minutes
16:07:11 <RRSAgent> I have made the request to generate https://www.w3.org/2026/08/27-webmachinelearning-minutes.html anssik
16:09:29 <anssik> s/to case/to care
16:12:22 <anssik> s/be give/we give
16:12:58 <anssik> s/if. you/if you
16:14:43 <anssik> RRSAgent, draft minutes
16:14:45 <RRSAgent> I have made the request to generate https://www.w3.org/2026/08/27-webmachinelearning-minutes.html anssik
18:07:15 <Zakim> Zakim has left #webmachinelearning