14:47:38 <RRSAgent> RRSAgent has joined #webmachinelearning
14:47:43 <RRSAgent> logging to https://www.w3.org/2026/02/12-webmachinelearning-irc
14:47:43 <Zakim> RRSAgent, make logs Public
14:47:44 <Zakim> please title this meeting ("meeting: ..."), anssik
14:47:44 <anssik> Meeting: WebML WG Teleconference – 12 February 2026
14:47:48 <anssik> Chair: Anssi
14:47:53 <anssik> Agenda: https://github.com/webmachinelearning/meetings/blob/main/telcons/2026-02-12-wg-agenda.md
14:48:01 <anssik> Scribe: Anssi
14:48:12 <anssik> scribeNick: anssik
14:48:22 <anssik> Present+ Anssi_Kostiainen
14:48:32 <anssik> RRSAgent, draft minutes
14:48:33 <RRSAgent> I have made the request to generate https://www.w3.org/2026/02/12-webmachinelearning-minutes.html anssik
14:58:49 <anssik> Present+ Dwayne_Robinson
15:00:06 <DwayneR> DwayneR has joined #webmachinelearning
15:00:15 <anssik> Present+ Markus_Tavenrath
15:00:44 <anssik> Present+ Jonathan_Ding
15:00:56 <anssik> Regrets+ Tarek_Ziade
15:02:11 <anssik> Present+ Rafael_Cintron
15:02:46 <anssik> Present+ Reilly_Grant
15:03:18 <mtavenrath> mtavenrath has joined #webmachinelearning
15:03:27 <handellm> handellm has joined #webmachinelearning
15:03:35 <RafaelCintron> RafaelCintron has joined #webmachinelearning
15:03:56 <anssik> Present+ Markus_Handell
15:04:35 <anssik> Present+ Christian_Liebel
15:05:03 <anssik> Topic: Incubations
15:05:15 <anssik> Subtopic: Dynamic AI Offloading Protocol
15:05:20 <anssik> gb, this is webmachinelearning/daop
15:05:20 <gb> anssik, OK.
15:05:36 <anssik> Anssi: at our last meeting we resolved to create an explainer for Dynamic AI Offloading Protocol (DAOP) and initiate prototyping
15:05:52 <anssik> -> https://www.w3.org/2026/01/15-webmachinelearning-minutes.html#89af
15:06:00 <ningxin> ningxin has joined #webmachinelearning
15:06:00 <anssik> ... thank you to MarkusH for providing feedback from Google Meet product perspective
15:06:15 <anssik> Anssi: today, Jonathan will briefly present the explainer and prototype that he has put together (thanks!)
15:06:27 <anssik> ... you'll find the content that is presented by Jonathan from PR #1
15:06:28 <gb> https://github.com/webmachinelearning/daop/pull/1 -> Pull Request 1 Add DAOP explainer and estimateQoS() illustration with background blur demo (by jonathanding)
15:06:38 <anssik> -> Explainer (PR review) https://github.com/webmachinelearning/daop/blob/feature-initial-content/README.md
15:06:47 <anssik> -> Explainer (PR preview) https://github.com/webmachinelearning/daop/blob/feature-initial-content/README.md
15:06:54 <anssik> -> Reference implementation for illustration purposes (PR preview) https://github.com/webmachinelearning/daop/tree/feature-initial-content/daop-illustration
15:07:07 <anssik> Anssi: after the presentation and demo, questions and comments are welcome
15:07:12 <anssik> ... and after the meeting, more detailed review feedback is welcome via PR #1
15:07:35 <Ehsan> Ehsan has joined #webmachinelearning
15:07:49 <anssik> Anssi: Jonathan, to start, please introduce the explainer
15:08:48 <anssik> Jonathan: this proposal extends the WebNN API
15:09:37 <anssik> ... DAOP proposal uses a model centric estimation, can be done at initialization time or dynamically during runtime
15:10:12 <anssik> ... the proposed estimateQoS() API extends the WebNN API
15:11:14 <anssik> ... this API uses performance tiers: "excellent", "good", "fair", "moderate", "slow", "very-slow", "poor"
15:11:42 <anssik> ... we propose a core API for performance negotiation
15:11:53 <anssik> ... to maximize the benefits of DAOP, the underlying WebNN specification should support a weightless build mode
15:12:01 <anssik> ... we propose that WebNN builders be extended to allow:
15:12:14 <anssik> ... - weightless constants
15:12:18 <anssik> ... - lazy / explicit binding
15:13:12 <anssik> ... instead of performance tier, the local inference engine can respond whether it can meet the givenv performance requirement using the proposed meetsRequirement() API
15:13:19 <anssik> Present+ Ningxin_Hu
15:13:25 <anssik> Present+ Ehsan_Toreini
15:14:20 <anssik> Jonathan: for prototyping, we polyfilled WebNN and provided a demo
15:14:45 <anssik> [Jonathan sharing DAOP demo]
15:16:35 <anssik> Jonathan: background blur demo uses lazy binding for performance benefit and avoids fingerprinting by using performance tiers
15:17:09 <anssik> ... micro-benchmarks can be re-run in the background also
15:18:03 <anssik> ... the explainer welcomes your comments and feedback in PR #1
15:18:03 <gb> https://github.com/webmachinelearning/daop/pull/1 -> Pull Request 1 Add DAOP explainer and estimateQoS() illustration with background blur demo (by jonathanding)
15:18:29 <anssik> ... we chose the group's preferred approach, and rejected the alternative design
15:18:32 <anssik> q?
15:19:46 <RafaelCintron> q+
15:19:50 <anssik> ack RafaelCintron
15:20:28 <anssik> Rafael: if the implementation runs the benchmark locally, why browses are better qualified to run the benchmark?
15:21:47 <anssik> Jonathan: browser has knowledge of the underlying platform and its capabilities, so can provide a better offline estimation, also to estimate the QoS holds over a longer period of time
15:22:45 <anssik> Rafael: thanks, that clarifies the question, sometimes the browser indeed knows better on which devices certain ops run
15:23:14 <anssik> Jonathan: we can provide a library abstraction on top of the proposed QoS estimation API
15:23:43 <anssik> ... the ultimate goal is to solve this problem for both web and native
15:23:47 <mtavenrath> q+
15:24:55 <anssik> Rafael: the guarantee of QoS works for games, but the web browser may be more complex as it has a number of tabs doing other tasks whose performance profile may change over time
15:24:56 <anssik> q?
15:26:14 <anssik> ... a learning from the Prompt API estimation was we don't want to run the user's machine to the ground when the API is called
15:27:16 <anssik> ... I think estimation sometimes is the best we can do, it can just be challenging and how to maintain it over time
15:27:18 <anssik> q?
15:27:31 <anssik> ack mtavenrath
15:27:59 <anssik> MarkusT: for estimating ops themselves may be problematic due to fusion that can happen
15:28:06 <reillyg> q+ to mention compatibility detection.
15:28:25 <handellm> handellm has joined #webmachinelearning
15:28:42 <anssik> ... there may be background operations that impact the estimation
15:29:26 <anssik> ... I'd be considered if performance metrics measured by the frameworks would be shared across websites to avoid super cookie-style mechanism
15:29:37 <anssik> q?
15:30:14 <anssik> ... generally I see this is useful as a library
15:30:50 <anssik> Jonathan: eventually we need native support to understand fusion, caching etc. optimizations only native knows about
15:30:52 <anssik> q?
15:31:07 <anssik> ack reillyg
15:31:07 <Zakim> reillyg, you wanted to mention compatibility detection.
15:31:44 <anssik> Reilly: I think agree with MarkusT that this is probably not dangerous from fingerprinting perspective
15:33:00 <anssik> ... for compatibility detection of models, developers could try to build models of different shapes, the challenge in that is the requirement to download all the models, but with weightless graph shape can do initial testing, interested to see what opportunities that provides
15:33:01 <anssik> q?
15:33:50 <anssik> q?
15:34:10 <mtavenrath> One could build a performance database and use WebGL to query the GPU string...
15:34:17 <anssik> Subtopic: Last Week in Community Group
15:34:21 <anssik> gb, this is webmachinelearning/webmcp
15:34:22 <gb> anssik, OK.
15:34:27 <anssik> -> https://github.com/webmachinelearning/meetings/blob/main/telcons/2026-02-05-cg-agenda.md
15:34:31 <anssik> -> https://www.w3.org/2026/02/05-webmachinelearning-minutes.html
15:34:38 <anssik> Anssi: the group established the WebMCP spec draft and landed the initial WebIDL in PR #75
15:34:38 <gb> https://github.com/webmachinelearning/webmcp/pull/75 -> MERGED Pull Request 75 Add WebIDL and descriptions to spec draft (by bwalderman)
15:34:42 <anssik> -> WebMCP latest draft https://webmachinelearning.github.io/webmcp/
15:35:05 <anssik> Anssi: the group initiated work on WebMCP accessibility consideration, the group believes there's a great potential to improve accessibility with this tech, new a11y experts joined to contribute to this effort
15:35:12 <anssik> gb, this is webmachinelearning/translation-api
15:35:12 <gb> anssik, OK.
15:35:29 <anssik> Anssi: for Translator and Language Detector APIs, there's been an active debase on input formats in Mozilla's standards position repo, see spec issue #60
15:35:29 <gb> https://github.com/webmachinelearning/translation-api/issues/60 -> Issue 60 Input formats for translation and language detection (by tomayac)
15:35:35 <anssik> gb, this is webmachinelearning/prompt-api
15:35:36 <gb> anssik, OK.
15:35:45 <anssik> Anssi: for Prompt API, we discussed how to address the TAG review feedback
15:35:56 <anssik> q?
15:36:03 <anssik> Topic: Web Neural Network API
15:36:07 <anssik> gb, this is webmachinelearning/webnn
15:36:07 <gb> anssik, OK.
15:36:12 <anssik> Subtopic: Origin Trial launch
15:36:23 <anssik> Anssi: I want to take a moment to celebrate WebNN Origin Trial launch on Chrome and Edge 146
15:36:32 <anssik> ... WebNN Origin Trial journey started in Chrome Beta on 11 Feb and hits Stable on 10 Mar
15:36:45 <anssik> ... and for Edge, Beta on 19 Feb and Stable on 12 Mar
15:37:00 <anssik> ... this significant implementation milestone follows hot on the heels of the WebNN API spec CRS milestone 22 Jan 2026
15:37:04 <anssik> ... congrats to the group for this major achievement!
15:37:18 <anssik> ... in concert with OT, the group also launched WebNN Docs, a vendor-neutral resource that assists early adopters on their OT journey
15:37:25 <anssik> -> WebNN Docs: Origin Trials Registration
15:37:25 <anssik>  https://webnn.io/en/learn/get-started/ot_registration
15:37:33 <anssik> -> WebNN Docs: Tutorials https://webnn.io/en/learn/tutorials
15:37:42 <anssik> Anssi: huge thanks to Belem for kicking off this docs resource!
15:37:50 <anssik> ... the entire webnn.io docs site is a community resource and you can contribute via GitHub:
15:37:54 <anssik> -> https://github.com/webmachinelearning/webnn-docs
15:38:06 <anssik> Anssi: any other OT resources to bring to the group's attention?
15:38:28 <ningxin> q+
15:38:28 <anssik> q?
15:38:35 <anssik> ack ningxin
15:39:31 <anssik> Ningxin: I wanted to share that webnn-samples already added OT token so feel free to use the samples without the flag
15:40:04 <RafaelCintron> https://aks.ms/webnn
15:40:08 <anssik> Rafael: I did the same for Edge
15:40:12 <RafaelCintron> https://aka.ms/webnn
15:40:42 <anssik> Subtopic: Power preferences and the fallback concept
15:40:47 <anssik> Anssi: issue #911
15:40:47 <gb> https://github.com/webmachinelearning/webnn/issues/911 -> Issue 911 accelerated should be prior to powerPreference for device selection (by mingmingtasd) [device selection]
15:40:51 <anssik> Anssi: two things:
15:41:10 <anssik> Anssi: first, I'd like to see if we can agree on the proposed spec change to expand powerPreference enum with a new value "fallback" (prior "no-acceleration")
15:41:22 <anssik> ... as discussed last time, "fallback" as a name would be more future-proof and conceptually aligned with WebGPU, thanks MikeW for the suggestion
15:41:29 <anssik> ... I feel this naming consideration is more than just bikeshedding
15:41:44 <anssik> q?
15:42:08 <anssik> ... second, I'd like to understand if we need any spec language changes to help clarify that these hints are non-normative
15:42:25 <anssik> ... we already avoid using any RFC 2119 MUSTs and SHOULDs in MLContextOptions spec section
15:42:29 <anssik> -> https://www.w3.org/TR/webnn/#api-mlcontextoptions
15:42:40 <anssik> Anssi: perhaps we could add a green Note box to highlight this fact
15:43:27 <anssik> Rafael: in both WebGPU/GL power preferences are hints
15:43:49 <RafaelCintron> https://browserleaks.com/webgl and https://browserleaks.com/webgpu
15:44:52 <reillyg> https://gpuweb.github.io/gpuweb/#dom-gpurequestadapteroptions-powerpreference
15:45:22 <anssik> Anssi: it seems we agree "fallback" is a good name for a new powerPreference and these hints are non-normative, and maybe we want to lift some language from WebGPU/GL for WebNN
15:45:57 <reillyg> q+
15:45:59 <anssik> q?
15:46:00 <anssik> ack reillyg
15:46:24 <anssik> Reilly: I think we're at the stage where PR with a specific proposal would be a reasonable next step
15:47:09 <anssik> q?
15:47:22 <anssik> Subtopic: Dynamic dimensions
15:47:26 <anssik> Anssi: issue #883
15:47:27 <gb> https://github.com/webmachinelearning/webnn/issues/883 -> Issue 883 Support flexible input sizes (by huningxin) [feature request] [operator specific] [Agenda+]
15:47:53 <anssik> Anssi: I'd like to review the latest MLDynamicDimension proposal and discuss any new information from prototyping
15:48:18 <anssik> ... Ningxin and Bin verified that the latest Chromium prototype of this feature is able to run TinyLlama 1.1B model with dynamic input size
15:48:43 <anssik> ... TinyLlama 1.1B is optimized for use on devices with restricted memory budget and computational capabilities and enables use cases such as real-time machine translation
15:48:48 <anssik> -> https://www.w3.org/TR/webnn/#usecase-translation
15:48:56 <anssik> Anssi: this bigger model is thus well-suited to demonstrate the MLDynamicDimension feature
15:49:14 <anssik> Anssi: prototyping experience suggested two additional IDL changes, change type of expand.newShape and reshape.newShape from unsigned long to MLDimension
15:50:39 <anssik> Ningxin: we can full chat demo with this feature, in my comments I gave an example how the developer can build this model with dynamic input and set the concrete dimensions for the input at different stages, prefill stage per sequence length, and in the decoding loop increase the total sequence length during the loop while still using the same model
15:51:16 <anssik> ... at last TPAC we shared static shape limitation, where the developers has to use two separate models, for prefill and decode, this increases memory footprint and inference time
15:51:28 <anssik> ... now with dynamic shape support we can use just one model
15:51:58 <anssik> ... for reshape and expand we observed changes are required:
15:52:05 <anssik> ```diff
15:52:05 <anssik> - MLOperand expand(MLOperand input, sequence<unsigned long> newShape, optional MLOperatorOptions options = {});
15:52:05 <anssik> + MLOperand expand(MLOperand input, sequence<MLDimension> newShape, optional MLOperatorOptions options = {});
15:52:06 <anssik> 
15:52:06 <anssik> - MLOperand reshape(MLOperand input, sequence<unsigned long> newShape, optional MLOperatorOptions options = {});
15:52:06 <anssik> + MLOperand reshape(MLOperand input, sequence<MLDimension> newShape, optional MLOperatorOptions options = {});
15:52:07 <anssik> ```
15:53:07 <anssik> Ningxin: in the prototype we can expand known dynamic dimension, we introduced known dynamic dimension that is defined by builder.input
15:54:11 <anssik> ... you can reshape or expand a shape with dynamic dimensions without introducing a new concept
15:54:25 <anssik> s/```diff//
15:55:07 <anssik> Ningxin: we believe we can also support Qwen and other LLMs with this proposed change
15:55:29 <anssik> ... Tarek's comments in rustnn are very helpful
15:55:29 <anssik> q?
15:55:53 <anssik> Ningxin: to conclude, by using one model for both prefill and decode stages, the session creation time and memory footprint are reduced significantly
15:56:50 <anssik> Anssi: Tarek is implementing and Markus is reviewing the the dynamic shape feature in rustnn, you can follow the plan and implementation progress
15:56:56 <anssik> -> https://github.com/rustnn/rustnn/issues/15
15:56:57 <gb> https://github.com/rustnn/rustnn/issues/15 -> Issue 15 Support flexible input sizes (by tarekziade)
15:57:01 <anssik> -> https://github.com/rustnn/rustnn/pull/16
15:57:01 <gb> https://github.com/rustnn/rustnn/pull/16 -> Pull Request 16 Add dynamic Dimension support (flexible input shapes) (by tarekziade)
15:57:10 <anssik> MarkusT: I will convey Tarek's feedback
15:57:45 <anssik> ... in summary the changes seemed to go well for rustnn
15:58:55 <anssik> q?
15:59:15 <reillyg> q+
15:59:18 <anssik> ack reillyg
15:59:56 <anssik> Reilly: I support the group advancing to a spec PR for this feature
16:00:44 <mtavenrath> q+
16:01:07 <anssik> Markus: for the API change, do we support both the versions, reshape and expand?
16:01:08 <ningxin> q+
16:01:11 <anssik> ack ningxin
16:01:49 <anssik> Ningxin: the current proposal is a union, the IDL changes but supports both static and dynamic use cases can run all the static samples
16:02:42 <reillyg> +1
16:02:48 <RafaelCintron> +1
16:02:48 <mtavenrath> +1
16:02:49 <anssik> RESOLUTION: Create a spec PR for MLDynamicDimension to support models that may require flexible input sizes, supports both static and dynamic use cases.
16:03:11 <anssik> Topic: Next WG Teleconference 12 March 2026
16:03:20 <anssik> Anssi: given many group participants are celebrating the Spring Festival, I will skip over the 26 Feb WG teleconference
16:03:26 <anssik> ... the next WG teleconference will be 12 March 2026
16:03:46 <anssik> ... I want to wish everyone who is celebrating Spring Festival very relaxing vacation
16:03:58 <anssik> ... for 2026 the animal is Horse representing energy, speed, freedom, and success
16:04:14 <anssik> ... I like this as it reflects this group and its timely launch of WebNN OT in 2026
16:04:23 <anssik> ... we will be back with even more energy and speed after this mini-break!
16:04:24 <anssik> q?
16:04:27 <ningxin> thanks Anssi!
16:05:04 <anssik> RRSAgent, draft minutes
16:05:05 <RRSAgent> I have made the request to generate https://www.w3.org/2026/02/12-webmachinelearning-minutes.html anssik
16:06:06 <anssik> RRSAgent, draft minutes
16:06:07 <RRSAgent> I have made the request to generate https://www.w3.org/2026/02/12-webmachinelearning-minutes.html anssik
16:08:19 <anssik> s/```//
16:08:21 <anssik> RRSAgent, draft minutes
16:08:22 <RRSAgent> I have made the request to generate https://www.w3.org/2026/02/12-webmachinelearning-minutes.html anssik
16:19:20 <anssik> s|-> Explainer (PR review) https://github.com/webmachinelearning/daop/blob/feature-initial-content/README.md||
16:19:22 <anssik> RRSAgent, draft minutes
16:19:23 <RRSAgent> I have made the request to generate https://www.w3.org/2026/02/12-webmachinelearning-minutes.html anssik
16:22:04 <anssik> s/givenv/given
16:22:38 <anssik> s/browses/browsers
16:23:11 <anssik> s/to estimate the QoS holds/can estimate the QoS that holds
16:24:22 <anssik> s/be considered/be concerned
16:25:01 <anssik> s/I think agree/I agree
16:27:32 <anssik> s|https://aks.ms/webnn||
16:35:11 <anssik> s/we can full/we can implement full
16:39:13 <anssik> RRSAgent, draft minutes
16:39:14 <RRSAgent> I have made the request to generate https://www.w3.org/2026/02/12-webmachinelearning-minutes.html anssik
17:48:49 <gb> gb has joined #webmachinelearning
18:26:37 <Zakim> Zakim has left #webmachinelearning