14:57:06 <RRSAgent> RRSAgent has joined #webmachinelearning
14:57:06 <RRSAgent> logging to https://www.w3.org/2022/11/17-webmachinelearning-irc
14:57:08 <Zakim> RRSAgent, make logs Public
14:57:09 <Zakim> please title this meeting ("meeting: ..."), anssik
14:57:19 <anssik> Meeting: WebML WG Teleconference – 17 November 2022
14:57:24 <anssik> Chair: Anssi
14:57:28 <anssik> Agenda: https://github.com/webmachinelearning/meetings/blob/main/telcons/2022-11-17-wg-agenda.md
14:57:32 <anssik> Scribe: Anssi
14:57:36 <anssik> scribeNick: anssik
14:59:36 <anssik> Present+ Anssi_Kostiainen
14:59:42 <anssik> Regrets+ Dominique_Hazael-Massieux
14:59:47 <anssik> RRSAgent, draft minutes
14:59:47 <RRSAgent> I have made the request to generate https://www.w3.org/2022/11/17-webmachinelearning-minutes.html anssik
15:01:07 <anssik> Present+ Jonathan_Bingham
15:01:53 <anssik> Present+ Dwayne_Robinson
15:01:56 <jonathan> jonathan has joined #webmachinelearning
15:01:59 <anssik> Present+ Zoltan_Kis
15:02:06 <ningxin_hu> ningxin_hu has joined #webmachinelearning
15:02:12 <anssik> Present+ Ningxin_Hu
15:02:25 <anssik> Present+ Rafael_Cintron
15:02:25 <jonathan> special guest: Eugene Burmako, who will co-present with Stella
15:02:43 <anssik> Present+ Eugene_Burmako
15:03:03 <RafaelCintron> RafaelCintron has joined #webmachinelearning
15:05:47 <burmako> burmako has joined #webmachinelearning
15:05:52 <dwayner> dwayner has joined #webmachinelearning
15:07:05 <anssik> Topic: XLA, OpenXLA Project, StableHLO
15:07:24 <anssik> Slideset: https://lists.w3.org/Archives/Public/www-archive/2022Nov/att-0001/XLA-Stella-Laurenzo.pdf
15:08:11 <anssik> [slide 1]
15:08:41 <anssik> s/Stella-Laurenzo/Eugene_Burmako
15:08:48 <anssik> [slide 2]
15:09:13 <anssik> Present+ Chai_Chaoweeraprasit
15:10:29 <anssik> [slide 3]
15:12:00 <anssik> RRSAgent, draft minutes
15:12:00 <RRSAgent> I have made the request to generate https://www.w3.org/2022/11/17-webmachinelearning-minutes.html anssik
15:17:43 <anssik> [slide 4]
15:19:43 <chai> chai has joined #webmachinelearning
15:21:11 <anssik> [slide 5]
15:23:48 <anssik> [slide 6]
15:24:41 <anssik> [slide 7]
15:28:05 <anssik> [slide 8]
15:29:41 <anssik> [slide 9]
15:30:11 <anssik> [slide 10]
15:31:46 <anssik> q+ to ask about StableHLO adoption by compiler projects
15:32:07 <anssik> ack anssik
15:32:07 <Zakim> anssik, you wanted to ask about StableHLO adoption by compiler projects
15:32:59 <anssik> anssik: StableHLO adoption in XLA compiler and IREE, which compiler is the leading adopter?
15:33:32 <anssik> Eugene: re main target for StableHLO, we provide equal support for these two compilers
15:34:22 <anssik> ... both take StableHLO input and can reflect all the functionality in XLA, cannot speak for IREE personally, but we're closely working with that team on new interesting applications
15:34:39 <anssik> ... Stella could confirm we can cover ~95% of use case
15:35:22 <zkis> zkis has joined #webmachinelearning
15:36:04 <anssik> anssik: do we still have an experimental compiler that interacts with StableHLO?
15:36:43 <anssik> Eugene: I wish Stella could speak for this topic, IREE is addressing some of the most important emerging ML use cases
15:38:19 <anssik> ... consumers of StableHLO, there's also TFLite
15:39:08 <anssik> ... our aspirations go beyond support in Google initiated compiler projects
15:40:22 <anssik> Chai: we are aware of XLA, we did some work TF codebase
15:40:47 <anssik> ... on Msft side we haven't leveraged XLA HLO yet, but are familiar with it, moving to open governance is a great thing
15:41:09 <anssik> ... in the process of defining the WebNN op set we also spent time looking at HLO so we can translate to HLO
15:41:21 <anssik> ... when we start with a set of ops it must make sense elsewhere
15:41:41 <anssik> ... looking at HLO op set is very helpful to understand how a reduced op set maps to actual models
15:41:49 <anssik> ... thanks for your work on this
15:42:15 <anssik> Chai: one question re interop with GPU
15:42:41 <anssik> ... you mentioned XLA is used a lot inside Google, what is the model to interop with the GPU stack? E.g. WebGPU with Project Dawn, how is that supported?
15:43:00 <anssik> Eugene: currently HLO is focused on datacenter usages
15:43:25 <anssik> ... this is the most tested path, we have aspirations to expand beyond servers but cannot speak for details there
15:44:03 <anssik> Chai: Computer Vision running on the server usage, this will eventually touch the graphics stack, how does that work with HLO?
15:45:04 <anssik> Eugene: within Alphabet we support multiple frameworks, TF, JAX, PyTorch etc. if involved with XLA compiler, JAX operated on it, the framework has to produce HLO to feed into compiler
15:45:50 <zkis_> zkis_ has joined #webmachinelearning
15:45:57 <anssik> ... TF has graph IR, stored, loaded, then we have TFXLA bridge that translates the graphs and compiles to XLA
15:46:14 <anssik> ... a bunch of ops cannot be compiler to XLA
15:46:35 <anssik> ... PyTorch uses lazy tensor tracing based mechanism to transform python programs into HLO
15:47:39 <anssik> ... within the compiler we have target independent optimizations, starting with simple things to advanced optimizations
15:47:48 <anssik> ... GPU compiler is available publicly, now in TF will be a separate project soon
15:48:17 <anssik> ... XLA GPU architecture on the high level, we also rewrite this arch to use MLIR more and more
15:48:22 <Github> Github has joined #webmachinelearning
15:48:34 <anssik> ... MLIR is very influential tech we love here at Google
15:49:11 <anssik> ... to recap, we support many frameworks with the same interface
15:49:14 <anssik> q?
15:49:27 <ningxin_hu> q+
15:50:44 <chai> q+
15:51:10 <anssik> anssik: StableHLO spec stability, timeline?
15:51:20 <anssik> Eugene: targeting the first version of the spec EOY
15:51:32 <anssik> ... calling it v1.0 or v0.9
15:51:43 <anssik> ... feature complete EOY
15:52:02 <anssik> ... active work in progress, we have half of it specced and the rest in open issues
15:52:33 <anssik> ... not a perfect time to look at the spec right now, maybe beginning of 2023 is a good date for WebNN op set compatibility effort
15:52:41 <anssik> -> StableHLO Spec draft https://github.com/openxla/stablehlo/blob/main/docs/spec_draft.md
15:52:46 <anssik> -> StableHLO Spec open issues https://github.com/openxla/stablehlo/labels/Spec
15:52:52 <anssik> -> StableHLO Spec index of ops https://github.com/openxla/stablehlo/blob/main/docs/spec_draft.md#index-of-ops
15:52:57 <stlukey97> stlukey97 has joined #webmachinelearning
15:53:30 <anssik> Eugene: areas of major development after 1.0: sparsity,  uniform quantization
15:53:52 <anssik> ... aligning ops between WebNN and StableHLO is a great idea, happy to support there
15:54:45 <anssik> ningxin_hu: thanks Eugene, great presentation!
15:55:01 <anssik> ... two groups looking to coordinate on the op sets makes a lot of sense
15:55:19 <anssik> ... question re op sets, you mentioned op set categories, also usage on server and mobile
15:55:56 <anssik> ... my impression is some ops are not so useful on mobile, distribution on multiple nodes, do you plan to have profiles in StableHLO e.g. for server and other classes?
15:56:25 <anssik> Eugene: we are interested in having those profiles, haven't done a full study yet on this topic
15:56:40 <anssik> ... we have this as a 2023 goal
15:56:57 <anssik> ... I agree distribution ops have no usefulness on mobile
15:57:33 <anssik> ningxin_hu: another questions re compilers, you mentioned XLA and IREE compilers, for web usage we want to understand if these compilers support JIT?
15:57:51 <anssik> ... on-device compilation would be an interesting feature for us
15:58:08 <anssik> Eugene: in general non-server story currently is where we have less clarity, an area of active work
15:58:30 <anssik> ... starting with server, XLA compiler is a JIT compiler predomantly
15:58:43 <anssik> ... OTOH, IREE is ahead-of-time
15:59:22 <anssik> ... on HLO side, a limitation is no dynamic shapes, no unknown tensor sizes, in StableHLO we address this, if we want to do AOT compilation with dynamic tensor sizes it becomes feasible, that is what IREE does
15:59:38 <anssik> ... many practical programs are dynamically sized
15:59:49 <anssik> ... on mobile we look both AOT and JIT use cases
16:00:52 <anssik> ningxin_hu: question re TFLite, WebNN works with TFLite Wasm version, in your slides you mentioned StableHLO is consumer as a flatbuffer schema, also TFLite delegate could leverage StableHLO
16:01:09 <anssik> ... is TFLite a consumer of StableHLO?
16:01:33 <anssik> Eugene: what I shared on the slide is as much as I could share on behalf of the TFLite team, WIP, a bit too early to share the details
16:01:34 <anssik> q?
16:01:44 <anssik> ack ningxin_hu
16:01:51 <anssik> ack chai
16:02:08 <anssik> chai: Eugene you said you want to push this for mobile too
16:02:29 <anssik> ... WebNN is the op set for the web, so I think there's some synergy around mapping StableHLO on top WebNN for mobile use case
16:02:40 <anssik> ... this is an area of collaboration where we can connect the dots
16:04:16 <anssik> anssik: we may want to add OpenXLA in the coordination for this WG's charter
16:04:23 <anssik> Eugene: sounds great
16:04:30 <ningxin_hu> +1
16:05:38 <anssik> RRSAgent, draft minutes
16:05:38 <RRSAgent> I have made the request to generate https://www.w3.org/2022/11/17-webmachinelearning-minutes.html anssik
16:13:20 <burmako> Thank you everyone for the invitation to present! It was great to meet you all, and I'm looking forward to collaborating.
16:13:47 <burmako> For the meeting minutes, is there a way to propose some edits?
16:13:55 <burmako> And I'll also share the slides shortly.
16:14:04 <anssik> anssik: thank you Eugene for the presentation!
16:58:05 <burmako> burmako has joined #webmachinelearning
17:06:58 <burmako> Hi Anssi! Not sure if my previous message reached the channel (got what seems to be an error message), but I emailed you the slides and proposed edits. Thanks again!
17:28:15 <anssik> s/ and can reflect all the functionality in XLA,/StableHLO supports all the functionality of XLA;
17:30:00 <anssik> s/HLO to feed into compiler/HLO to feed into compiler, JAX's mapping to HLO is fairly straightforward.
17:30:27 <anssik> s/a bunch of ops cannot be compiler to XLA/a bunch of TF ops cannot be compiler to XLA
17:31:05 <anssik> s/sparsity,  uniform quantization/finalizing sparsity, quantization including uniform quantization and beyond, extensibility
17:31:49 <anssik> s/I agree distribution ops have no usefulness on mobile/I agree distribution ops have limited or no usefulness on mobile
17:32:26 <anssik> s/on HLO side, a limitation is no dynamic shapes/a limitation is no dynamic shapes
17:33:26 <anssik> s/no unknown tensor sizes/no unknown tensor sizes (bounded dynamic dimension sizes are supported)
17:33:34 <anssik> RRSAgent, draft minutes
17:33:34 <RRSAgent> I have made the request to generate https://www.w3.org/2022/11/17-webmachinelearning-minutes.html anssik
17:54:47 <anssik> anssik: edits to the minutes proposed by Eugene have been incorporated (see Diagnostics log below)
17:54:49 <anssik> RRSAgent, draft minutes
17:54:49 <RRSAgent> I have made the request to generate https://www.w3.org/2022/11/17-webmachinelearning-minutes.html anssik
18:30:35 <Zakim> Zakim has left #webmachinelearning
19:12:55 <zkis_> zkis_ has joined #webmachinelearning