README: fix package path (esmfold2_trimul_kernel), use from_pretrained(use_kernels=True), note build layout

README.md

@@ -25,33 +25,50 @@ visibility)`, reading its parameters (`norm_start`/`norm_mix`/`proj_bundle`/
 attribute names and forward signature** — that's the contract.
 
 ```python
-model = ESMFold2Model.from_pretrained("biohub/ESMFold2", dtype=torch.bfloat16).cuda()
-model.set_use_kernels(True)            # swaps in this kernel for the 124 trimul sites
+import torch
+from transformers import ESMFold2Model
+
+# use_kernels=True swaps in this kernel for the 124 trimul sites (CUDA + inference).
+model = ESMFold2Model.from_pretrained(
+    "biohub/ESMFold2", dtype=torch.bfloat16, device_map="cuda", use_kernels=True
+).eval()
 out = model.infer_protein(seq)
 ```
 
 ## Layout
 
+The package name must match `kernels`' repo-derived name
+(`repo_id.split("/")[-1].replace("-", "_")`), i.e. **`esmfold2_trimul_kernel`** for the
+repo `…/esmfold2-trimul-kernel`, and `build.toml`'s `[general] name` must match it too.
+`kernels.get_kernel` loads from `build/torch-universal/` (not `torch-ext/`).
+
 ```
-build.toml                              # kernel-builder config (universal/Triton)
-flake.nix                               # kernel-builder entry (verify vs current version)
-torch-ext/esmfold2_trimul/
-  __init__.py                           # exports the layer + the functional entry
-  layers.py                             # ESMFold2TriangleMultiplication (the Hub layer)
-  trimul_with_residual.py               # kernel entrypoint
-  fused_dual_gemm.py                    # helper: gated dual GEMM
-  fused_ln_residual.py                  # helper: LN + transpose / residual-link epilogues
-  trimul_einsum_triton.py               # helper: batched triangular einsum
+build.toml                                       # kernel-builder config (universal/Triton)
+flake.nix                                        # kernel-builder entry (verify vs current version)
+torch-ext/esmfold2_trimul_kernel/                # source (read by kernel-builder)
+  __init__.py                                    # exports the layer + the functional entry
+  layers.py                                      # ESMFold2TriangleMultiplication (the Hub layer)
+  trimul_with_residual.py                        # kernel entrypoint
+  fused_dual_gemm.py                             # helper: gated dual GEMM
+  fused_ln_residual.py                           # helper: LN + transpose / residual-link epilogues
+  trimul_einsum_triton.py                        # helper: batched triangular einsum
+build/torch-universal/esmfold2_trimul_kernel/    # loaded by kernels.get_kernel (same files)
 ```
 
 ## Build & publish
 
+The `build/torch-universal/` dir checked in here is a hand-built universal layout
+(the Triton package copied in — no compile step), which is sufficient for
+`kernels.get_kernel`. To regenerate it properly with kernel-builder:
+
 ```bash
 nix build .#bundle      # or: kernel-builder build  (see kernel-builder docs)
-# then push the built repo to the Hub and set repo_id in transformers'
-# hub_kernels.py (currently the placeholder biohub/esmfold2-trimul-kernel).
 ```
 
+Mapped from transformers in `integrations/hub_kernels.py` under the layer name
+`ESMFold2TriangleMultiplication`. Currently `repo_id = Rocketknight1/esmfold2-trimul-kernel`
+(testing); move it to a `kernels-community` org and update `repo_id` before merging.
+
 ## Validation
 
 Swapped into all 124 `TriangleMultiplicativeBlock` instances of the real model