docs(upstream): receipt of 6 PRs filed 2026-05-04 by apstenku123

Files documenting the actual PRs we just opened upstream:

- PR #1: ml-explore/mlx#3476 — from_dlpack Metal-aware consumer (against main, clean)
- PR #2: apache/tvm#19504 — TVM_METAL_STORAGE_MODE env opt-in (against main, clean)
- PR #3: tile-ai/tilelang#2139 — mixed-dtype T.gemm via scalar fallback (stacks on PR #2130)
- PR #4: tile-ai/tilelang#2140 — FP8-input T.gemm scalar fallback routing (stacks on PR #2130)
- PR #5: tile-ai/tilelang#2141 — T.Pipelined num_stages>1 3D buffer fix (stacks on PR #2130)
- PR #6: tile-ai/tilelang#2142 — T.fp8_scaled_matmul DSL intrinsic (stacks on PR #2130)

Deferred (split into companion PRs needed): tilelang_metal_fp8 and
tilelang_metal_fp8_vector each touch both tilelang supermodule and the
TileLang/tvm vendored submodule. These need 2 PRs each — one to
tile-ai/tilelang, one to TileLang/tvm — separate filing round.

PRs #3-#6 are independent of each other; each branches directly from
jorgecurious/tilelang:metal-gemm-upstream-rebase HEAD 971c17b, so they
can be reviewed in any order. They DO depend on the upstream 4-PR Apple
Metal landing chain (#1869, #2118, #2121, #2130) merging first; if any
of those land separately, ours can be retargeted at main.

Author: apstenku123

docs/upstream/_filed_prs_2026_05_04.md

@@ -0,0 +1,69 @@
+# Filed upstream PRs — 2026-05-04
+
+GitHub user: `apstenku123` (David Gornshtein, davidgornshtein@gmail.com)
+
+## Filed (6 of 9 artifacts)
+
+| # | Repo | PR | Title | Status |
+|---|---|---|---|---|
+| 1 | `ml-explore/mlx` | [#3476](https://github.com/ml-explore/mlx/pull/3476) | [Python] Add mx.from_dlpack(obj) Metal-aware consumer | filed against `main` |
+| 2 | `apache/tvm` | [#19504](https://github.com/apache/tvm/pull/19504) | [Runtime][Metal] add TVM_METAL_STORAGE_MODE env opt-in for Shared/Managed buffers | filed against `main` |
+| 3 | `tile-ai/tilelang` | [#2139](https://github.com/tile-ai/tilelang/pull/2139) | [Metal] allow mixed-dtype T.gemm via scalar fallback | filed against `main`, **stacks on PR #2130** |
+| 4 | `tile-ai/tilelang` | [#2140](https://github.com/tile-ai/tilelang/pull/2140) | [Metal] route FP8-input T.gemm to scalar fallback | filed against `main`, **stacks on PR #2130** |
+| 5 | `tile-ai/tilelang` | [#2141](https://github.com/tile-ai/tilelang/pull/2141) | [Metal] thread stage dim through T.access_ptr for T.Pipelined num_stages>1 | filed against `main`, **stacks on PR #2130** |
+| 6 | `tile-ai/tilelang` | [#2142](https://github.com/tile-ai/tilelang/pull/2142) | tilelang: T.fp8_scaled_matmul DSL intrinsic + Metal lowering | filed against `main`, **stacks on PR #2130** |
+
+## Deferred (split needed: tilelang supermodule + TileLang/tvm submodule)
+
+These two patches modify both `src/target/codegen_metal.{cc,h}` (tilelang
+supermodule) and `3rdparty/tvm/src/target/source/codegen_metal.{cc,h}`
+(vendored TileLang/tvm submodule). They cannot be filed as a single
+tilelang PR because the supermodule and submodule have separate review
+chains. Split into companion PRs needed:
+
+| Artifact | Tilelang side | TileLang/tvm side |
+|---|---|---|
+| `tilelang_metal_fp8` | `src/target/codegen_metal.{cc,h}` (storage-only FP8 helpers) | `3rdparty/tvm/src/target/source/codegen_metal.{cc,h}` (mirror) |
+| `tilelang_metal_fp8_vector` | `src/target/codegen_metal.{cc,h}` (vector cast lanes 2/3/4) | `3rdparty/tvm/src/target/source/codegen_metal.{cc,h}` (mirror) |
+
+To file: each becomes 2 PRs:
+- One to `tile-ai/tilelang` (supermodule half)
+- One to `TileLang/tvm` (where the vendored TVM fork lives, per `.gitmodules:url = https://github.com/TileLang/tvm`)
+
+`tilelang_metal_shared_dyn` is the no-op investigation artifact, not a
+code PR — stays as `docs/` evidence only.
+
+## Stacking dependency on jorgecurious's branch
+
+PRs #3-#6 explicitly state in their bodies they stack on `jorgecurious/tilelang:metal-gemm-upstream-rebase` (PR #2130) at HEAD `971c17b`. That branch in turn stacks on:
+
+- [PR #1869 (oraluben)](https://github.com/tile-ai/tilelang/pull/1869) — base Metal GEMM landing
+- [PR #2118 (cklxx)](https://github.com/tile-ai/tilelang/pull/2118) — Metal scalar fallback for T.gemm
+- [PR #2121 (SiriusNEO)](https://github.com/tile-ai/tilelang/pull/2121) — multi-backend CodeGen refactor
+- [PR #2130 (jorgecurious)](https://github.com/tile-ai/tilelang/pull/2130) — metal-gemm-upstream-rebase
+
+All four are currently OPEN at time of filing (2026-05-04). When upstream merges that 4-PR chain into `tile-ai/tilelang:main`, our PRs #2139-#2142 become independently reviewable against main.
+
+## What each PR depends on (textually, for stacked review)
+
+```
+                ┌─────────────────────────────────────────────────────┐
+                │  Apple Metal landing chain (existing OPEN PRs)      │
+                │  #1869 → #2118 → #2121 → #2130                       │
+                └─────────────────────────────────────────────────────┘
+                                          │
+                                          ▼
+                ┌─────────────────────────────────────────────────────┐
+                │  Our 4 filed PRs (independent of each other)        │
+                │  #2139 mixed-dtype                                   │
+                │  #2140 FP8 gemm software path                        │
+                │  #2141 pipelined 3D buffer                           │
+                │  #2142 T.fp8_scaled_matmul                           │
+                └─────────────────────────────────────────────────────┘
+```
+
+(The 4 are independent — they all branch from #2130 directly, not from each other. Independent review.)
+
+## Path C blocker tracker
+
+For future Path C unblocks see `docs/upstream/_path_c_blockers_tracker.md`. 5 future patches identified (A: 32×32 pipelined extension; B: fused FP8 scaled scheduler; C: e8m0 blockscaled DSL; D: simdgroup_reduce primitive; E: chunked bwd language extension).