Merge OpenAI Triton commit d183197 (#4782)

This PR change the Triton base from
87f5aa4
to
d183197
(Jul 16).
Pass rate: 98.62%

Please do not squash and merge this PR.

Author: anmyachev

.github/workflows/integration-tests-nvidia.yml

@@ -13,6 +13,8 @@ jobs:
   integration-tests-nvidia:
     runs-on: ${{ matrix.runner }}
     timeout-minutes: 60
+    # Let A100 and H100 continue even if GB200 fails, as it's a bit flaky
+    continue-on-error: ${{ matrix.runner[0] == 'nvidia-gb200'}}
     strategy:
       matrix:
         runner: ${{ fromJson(inputs.matrix) }}

docs/meetups/for_moderators/README.md

@@ -0,0 +1,121 @@
+### How to run a Triton Community Meetup
+
+Contributors:  Bill Yoshimi, Areg Melikadamyan, Whitney Tsang, Ksharma Pawar
+
+Last updated: July 3, 2025
+
+Community meetups give the on-line community a chance to interact with each other and the Triton developers in a more face-to-face format vs slack chats.  Example topics covered during community meetups include:
+* Developers presenting updates on features they’re working on.
+* Developers asking community for feedback on new initiatives
+* Questions from community for developers
+* Questions about Triton strategy/direction.
+
+
+## Some logistics
+
+Community meetups occur once 8 weeks (usually during the first 1-2 weeks of a month).
+Reminders are sent out 2 weeks ahead of time
+
+Only companies that paid for corp Microsoft Teams access can create webinars.  Three folks who have done this (or have access in the past are):
+* Areg Melikadamyan
+* Whitney Tsang
+* Ksharma Pawar
+* Jian Hui
+
+Webinars are automatically recorded.  The person with corp access can upload the video to youtube after the webinar is finished.
+
+Only the person with corp access can open a webinar.  Even if you’re a registered speaker or MC, you’ll see
+
+<div align="center">
+  <img src="https://private-user-images.githubusercontent.com/85795580/459649517-42acf22a-6ae3-4c0f-a585-dffc25fabc33.png?jwt=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.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.UwzTbqMIsFw1Ex9t-5HY8A3KHzOnY6eqgqz2eJ0qTbE" alt="Microsoft Meeting waiting for meeting to start view">
+</div>
+
+
+before the meeting owner opens it up.
+
+During the meetup, take notes.
+
+Post the final notes on the Triton-lang website here: https://github.com/triton-lang/triton/tree/main/docs/meetups
+
+## How to run a community meetup
+
+1. Work with one of the folks above to create a Microsoft Teams webinar (occurring 6-8 weeks in the future).  Template:
+
+<pre>
+Title: “Triton Community Meetup (online)”
+External presenter: **“<your name>”**
+Co-organizer: **add organizers**
+    Date: **Add date**
+    Time: 10:00-11:00 PDT
+    Duration: 1 hr
+    Recurring meeting: link **(created by [email protected])**
+</pre>
+
+2. If you don’t have details about the meeting (e.g. meeting ID, passcode, phone number, etc.) you can login to the meeting, click on More -> Meeting Info and get data that way.
+
+<div align="center">
+  <img src="https://private-user-images.githubusercontent.com/85795580/459649518-f9a924df-2020-4dde-b8bb-e901a9298cef.png?jwt=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.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.6bgthh7pHDV-ooMMm1WsgbbLJH39nzk-DPre6GDjZnI" alt="Microsoft Meeting info pane">
+</div>
+
+3. Copy the invite generated from the meeting to [triton #general chat](https://app.slack.com/huddle/T01379XQ9FG/C013E22BPPC) on slack. Example text to use in slack:
+
+<pre>
+The next community meetup will be on <date> from 10am-11am PST. The meeting link is below. If anyone has agenda items to add for the meetup please reach out to me.
+
+Thanks,
+**<your name>**
+----
+Microsoft Teams Need help?
+Join the meeting now <- **change this**
+Meeting ID: 247 993 489 59 <- **change this**
+Passcode: Dg8TLJ <- **change this**
+Dial in by phone
++1 323-849-4874,,637393478# United States, Los Angeles <- **change this**
+Find a local number
+Phone conference ID: 637 393 478# <- **change this**
+</pre>
+
+4. Post the same invite to the [#triton channel on Discord GPU_MODE](https://discord.com/channels/1189498204333543425/1189607595451895918). You will need to join GPU_MODE to post to it.
+
+5. 1-2 Days before the meeting. Verify that someone with corp Microsoft Teams access will open the meeting up for you.
+6. Day before meeting, post reminders to slack and discord (reply to your original message):
+Reminder, this month's community meetup is tomorrow at 10am PST.
+
+<pre>
+Agenda:
+   Topic #1 <who>
+   Topic #2 <who>
+</pre>
+
+7. Day of meeting, login a little early and verify everything is working as expected.
+
+8. During the meeting, keep an eye on the comments section. Some folks might post questions for the speaker there and/or issues they're having with Teams.
+
+9. After the meeting has finished, work with the person with corp Microsoft Teams access to upload the recorded video to youtube.  Post the youtube link in [triton #general chat](https://app.slack.com/huddle/T01379XQ9FG/C013E22BPPC).
+
+If this is your first time using Microsoft Teams, work with the meeting creator to test out the UI (e.g. logging in, verifying your camera, audio work, verifying you can present your screen if using that functionality, play around with hand raising, play around with people/attendees/muting others, log off and log back in again.)
+
+## How to upload videos to Youtube
+
+1. Register to use youtube. That is, the top right symbol when you go to youtube.com should be your login.
+
+<div align="center">
+  <img src="https://private-user-images.githubusercontent.com/85795580/459649516-4bea87d7-1c6b-4331-92c1-e480b2f3d295.png?jwt=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.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.RUIK1GbqVtC8sfu_2ntb8to9FFQv7MfHrYbvNCfLCh8" alt="Youtube Login">
+</div>
+
+2. Click on https://www.youtube.com/feed/you
+3. Click on “+ Create” on top next to search box.
+4. Select the video you want to upload
+5. For Title use something like “Triton Community Meetup May 2025”
+6. No, it’s not made for kids
+7. No video elements
+8. Save or publish: “public”
+9. Make a copy of the video link so you can post it on slack and discord. (like: https://youtu.be/kJjBurkPn_8)
+
+
+## Past community meetups
+
+ | Date | Meet setup | Agenda & who | Recording |
+ | ---- | ---------- | ------------ | --------- |
+ | 2025-05-01 | [Link](https://tinyurl.com/mr397f6x) | Topic: what are plans for existing block pointer programming model? (Context: Intel GPU backend relies heavily on it and will need time to fully move to tensor descriptor programming model.) - Jianhui Li, Intel <br/> Topic: infrastructure for Triton performance tests - Sayce, Google<br/>Topic: what talks/tutorials/open discussions would you like to see at the 2025 Triton Developers’ Summit? How can we help? - Adnan Aziz, Meta <br/> Topic: what are plans for existing block pointer programming model? (Context: Intel GPU backend relies heavily on it and will need time to fully move to tensor descriptor programming model.) - Jianhui Li, Intel<br/>Topic: infrastructure for Triton performance tests - Sayce, Google<br/>Topic: what talks/tutorials/open discussions would you like to see at the 2025 Triton Developers’ Summit? How can we help? - Adnan Aziz, Meta </pre> | https://www.youtube.com/watch?v=W16BrXc5BYE |
+| 2025-07-09 |[Link](https://tinyurl.com/mus5wyax) | Topic: Gluon update - Jeff Niu, OpenAI <br/> Topic: Interest and requirements for a nightly performance regression suite - Simon Waters,  kernelize.ai<br/>Triton developer's summit update - Ofer Dekel, Microsoft | |

include/triton/Conversion/TritonGPUToLLVM/Utility.h

@@ -624,6 +624,25 @@ inline bool isCanonicalIndex(unsigned index, unsigned freeVarMask) {
 // group code isolated from above by invoking this function.
 void makeAllWarpGroupsIsolatedFromAbove(Operation *op);
 
+/// Converts ConverLayoutOp to llvm using padded pattern.
+/// This pattern adds unused memory locations after every rows of tensor fastest
+/// changing dimension:
+/// e0 e1 e2 e3 p p \
+/// e4 e5 e6 e7 p p \
+/// ...
+/// e e e e p p
+/// Dimension order is chosen in order to use wide output reads.
+///
+/// \param op operation to convert
+/// \param src llvm structure containing operation input
+/// \param targetInfo
+/// \param typeConverter
+/// \param rewriter
+/// \returns llvm structure containing converted output
+Value transferWithinBlockPadding(triton::gpu::ConvertLayoutOp op, Value src,
+                                 const TargetInfoBase &targetInfo,
+                                 const LLVMTypeConverter *typeConverter,
+                                 RewriterBase &rewriter);
 } // namespace mlir
 
 #endif

include/triton/Dialect/Gluon/IR/CMakeLists.txt

@@ -1,16 +1,17 @@
 set(MLIR_BINARY_DIR ${CMAKE_BINARY_DIR})
 
+set(LLVM_TARGET_DEFINITIONS GluonOps.td)
+mlir_tablegen(Ops.h.inc -gen-op-decls)
+mlir_tablegen(Ops.cpp.inc -gen-op-defs)
+add_mlir_doc(GluonOps GluonOps dialects/ -gen-op-doc)
+
 set(LLVM_TARGET_DEFINITIONS GluonDialect.td)
 mlir_tablegen(Dialect.h.inc -gen-dialect-decls -dialect=gluon)
 mlir_tablegen(Dialect.cpp.inc -gen-dialect-defs -dialect=gluon)
-mlir_tablegen(Ops.h.inc -gen-op-decls)
-mlir_tablegen(Ops.cpp.inc -gen-op-defs)
 add_mlir_doc(GluonDialect GluonDialect dialects/ -gen-dialect-doc)
-add_public_tablegen_target(GluonTableGen)
 
 set(LLVM_TARGET_DEFINITIONS GluonAttrDefs.td)
 mlir_tablegen(GluonAttrDefs.h.inc -gen-attrdef-decls)
 mlir_tablegen(GluonAttrDefs.cpp.inc -gen-attrdef-defs)
-mlir_tablegen(OpsEnums.h.inc -gen-enum-decls)
-mlir_tablegen(OpsEnums.cpp.inc -gen-enum-defs)
-add_public_tablegen_target(GluonAttrDefsIncGen)
+
+add_public_tablegen_target(GluonTableGen)

include/triton/Dialect/Gluon/IR/Dialect.h

@@ -6,3 +6,6 @@
 
 #define GET_ATTRDEF_CLASSES
 #include "triton/Dialect/Gluon/IR/GluonAttrDefs.h.inc"
+
+#define GET_OP_CLASSES
+#include "triton/Dialect/Gluon/IR/Ops.h.inc"

include/triton/Dialect/Gluon/IR/GluonOps.td

@@ -0,0 +1,32 @@
+#ifndef GLUON_OPS
+#define GLUON_OPS
+
+include "triton/Dialect/Gluon/IR/GluonDialect.td"
+include "triton/Dialect/Gluon/IR/GluonAttrDefs.td"
+include "triton/Dialect/Triton/IR/TritonInterfaces.td"
+include "triton/Dialect/Triton/IR/TritonTypes.td"
+
+class Gluon_Op<string mnemonic, list<Trait> traits = []> :
+    Op<Gluon_Dialect, mnemonic,
+       !listconcat(traits, [VerifyTensorLayoutsTrait])> {
+}
+
+def Gluon_SetAutoLayoutOp : Gluon_Op<"set_auto_layout",
+                                 [SameOperandsAndResultShape,
+                                  SameOperandsAndResultElementType]> {
+  let summary = "set auto encoding to a concrete encoding type";
+
+  let arguments = (ins TT_Tensor:$src);
+
+  let results = (outs TT_Tensor:$result);
+
+  let builders = [
+    OpBuilder<(ins "Attribute":$encoding, "Value":$value)>
+  ];
+
+  let hasVerifier = 1;
+
+  let assemblyFormat = "$src attr-dict `:` type($src) `->` type($result)";
+}
+
+#endif // GLUON_OPS

include/triton/Dialect/Triton/IR/TritonOps.td

@@ -432,6 +432,16 @@ def TT_SplatOp : TT_Op<"splat", [Pure,
     let hasFolder = 1;
 }
 
+def TT_UnsplatOp : TT_Op<"unsplat", [Pure,
+                                     DeclareOpInterfaceMethods<InferTypeOpInterface>]> {
+    let summary = "convert a tensor with a single element to a scalar";
+    let arguments = (ins TT_Tensor:$src);
+    let results = (outs TT_Type:$result);
+
+    let assemblyFormat = "$src attr-dict `:` type($src)";
+    let hasVerifier = 1;
+}
+
 def TT_ExpandDimsOp : TT_Op<"expand_dims", [Pure,
                                             DeclareOpInterfaceMethods<InferTypeOpInterface>,
                                             SameOperandsAndResultElementType]> {

include/triton/Dialect/TritonGPU/Transforms/Utility.h

@@ -260,11 +260,12 @@ void replaceUsesAndPropagateType(OpBuilder &builder, Operation *oldUse,
 
 /// Replace all uses of `old` with a local load from `alloc` unless the use is a
 /// `ttg.local_alloc` with a matching shared encoding, in which case the shared
-/// memory is forwarded directly into the use.
-void replaceUsesWithLocalLoad(
-    OpBuilder &builder, OpResult old,
-    TypedValue<triton::gpu::MemDescType> alloc,
-    TypedValue<triton::gpu::AsyncTokenType> token = {});
+/// memory is forwarded directly into the use. Returns the `ttg.local_load` if
+/// it created one.
+triton::gpu::LocalLoadOp
+replaceUsesWithLocalLoad(OpBuilder &builder, OpResult old,
+                         TypedValue<triton::gpu::MemDescType> alloc,
+                         TypedValue<triton::gpu::AsyncTokenType> token = {});
 
 // Return true if the value comes from a load or a block argument.
 // This will skip convert layouts and memdesc views.