miniworld-engine 2.3.9__tar.gz → 2.3.10__tar.gz

{miniworld_engine-2.3.9 → miniworld_engine-2.3.10}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: miniworld-engine
-Version: 2.3.9
+Version: 2.3.10
 Summary: MiniWorld GPU kernel development repo — Triton / CUDA kernels + benches for AF3-style ops (triangle multiplication, transition, attention, ...).
 License-Expression: MIT AND Apache-2.0 AND BSD-3-Clause
 Requires-Python: >=3.10
@@ -47,7 +47,7 @@ optimized in isolation: **hand-written CUDA where it exists, a Triton fallback o
 > blocks; terminal product repos assemble those into full models. Boundary rules live in
 > team-gm's `docs/ARCHITECTURE.md`.
 
-**Version 2.3.9** — CUDA + Triton only (CuTe/quack removed), torch 2.13 (CUDA 12.9 or 13.0), prebuilt H100 kernels,
+**Version 2.3.10** — CUDA + Triton only (CuTe/quack removed), torch 2.13 (CUDA 12.9 or 13.0), prebuilt H100 kernels,
 cuequivariance 0.12. [Changelog](docs/CHANGELOG.md) ·
 [attribution](licenses/THIRD_PARTY_NOTICES.md).
 
@@ -193,15 +193,18 @@ from the code.
 <!-- BEGIN GENERATED: hardware-support -->
 | arch | GPUs | kernels | backends |
 |---|---|---|---|
-| **sm80+** | A100, A5000, A6000, RTX 4090 | 132 | triton 126, cuda 6 |
+| **sm80+** | A100, A5000, A6000, RTX 4090 | 138 | triton 132, cuda 6 |
 | **sm90+** | H100 | 34 | cuda 30, triton 4 |
 <!-- END GENERATED: hardware-support -->
 
 **The minimum supported architecture is sm80.** The generated table above states each
 registered kernel's minimum architecture; it is not a numerical qualification for every
-listed GPU. Portable Triton paths and higher-architecture Triton/CUDA alternatives are
-selected by `modules/dispatch.py`. `implementation='triton'` explicitly selects the
-Triton backend; native CUDA variants can require a newer architecture.
+listed GPU. The public backends are `pytorch`, `miniworld`, `cuequivariance` and `anthropic`; Triton is
+not one of them. `miniworld` runs the engine's CUDA path where one serves the call; on a card without one
+it routes to the internal Triton kernel family only if the engine holds a record for that exact card
+(`autotune/manifests/`), else it warns and runs the PyTorch reference
+([docs/dispatch.md](docs/dispatch.md)). `settings.force_triton_kernels` is the internal A/B switch that
+forces the Triton family. Native CUDA variants can require a newer architecture.
 
 One extension is **not** in the table because it is not in the registry: `transition_b2b_cuda`,
 which the `Transition` module builds on demand, is compiled for `sm_90a` and fails to build on

{miniworld_engine-2.3.9 → miniworld_engine-2.3.10}/README.md

@@ -9,7 +9,7 @@ optimized in isolation: **hand-written CUDA where it exists, a Triton fallback o
 > blocks; terminal product repos assemble those into full models. Boundary rules live in
 > team-gm's `docs/ARCHITECTURE.md`.
 
-**Version 2.3.9** — CUDA + Triton only (CuTe/quack removed), torch 2.13 (CUDA 12.9 or 13.0), prebuilt H100 kernels,
+**Version 2.3.10** — CUDA + Triton only (CuTe/quack removed), torch 2.13 (CUDA 12.9 or 13.0), prebuilt H100 kernels,
 cuequivariance 0.12. [Changelog](docs/CHANGELOG.md) ·
 [attribution](licenses/THIRD_PARTY_NOTICES.md).
 
@@ -155,15 +155,18 @@ from the code.
 <!-- BEGIN GENERATED: hardware-support -->
 | arch | GPUs | kernels | backends |
 |---|---|---|---|
-| **sm80+** | A100, A5000, A6000, RTX 4090 | 132 | triton 126, cuda 6 |
+| **sm80+** | A100, A5000, A6000, RTX 4090 | 138 | triton 132, cuda 6 |
 | **sm90+** | H100 | 34 | cuda 30, triton 4 |
 <!-- END GENERATED: hardware-support -->
 
 **The minimum supported architecture is sm80.** The generated table above states each
 registered kernel's minimum architecture; it is not a numerical qualification for every
-listed GPU. Portable Triton paths and higher-architecture Triton/CUDA alternatives are
-selected by `modules/dispatch.py`. `implementation='triton'` explicitly selects the
-Triton backend; native CUDA variants can require a newer architecture.
+listed GPU. The public backends are `pytorch`, `miniworld`, `cuequivariance` and `anthropic`; Triton is
+not one of them. `miniworld` runs the engine's CUDA path where one serves the call; on a card without one
+it routes to the internal Triton kernel family only if the engine holds a record for that exact card
+(`autotune/manifests/`), else it warns and runs the PyTorch reference
+([docs/dispatch.md](docs/dispatch.md)). `settings.force_triton_kernels` is the internal A/B switch that
+forces the Triton family. Native CUDA variants can require a newer architecture.
 
 One extension is **not** in the table because it is not in the registry: `transition_b2b_cuda`,
 which the `Transition` module builds on demand, is compiled for `sm_90a` and fails to build on

{miniworld_engine-2.3.9 → miniworld_engine-2.3.10}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "miniworld-engine"
-version = "2.3.9"
+version = "2.3.10"
 requires-python = ">=3.10"
 readme = "README.md"
 description = "MiniWorld GPU kernel development repo — Triton / CUDA kernels + benches for AF3-style ops (triangle multiplication, transition, attention, ...)."

{miniworld_engine-2.3.9 → miniworld_engine-2.3.10}/src/miniworld_engine/architecture/exceptions.py

@@ -7,6 +7,7 @@ class ImplementationType(str, Enum):
     PYTORCH = "pytorch"
     MINIWORLD_ENGINE = "miniworld_engine"
     CUEQUIVARIANCE = "cuequivariance"
+    ANTHROPIC = "anthropic"
     # Source compatibility for old consumers; iteration uses the canonical member.
     MINIWORLD_KERNELS = MINIWORLD_ENGINE
 

{miniworld_engine-2.3.9 → miniworld_engine-2.3.10}/src/miniworld_engine/autotune/builder.py

@@ -318,7 +318,8 @@ def cases() -> list[Case]:
     )
 
     def IT(i):
-        return ImplementationType(i)
+        # "triton" is the A/B label of a build unit (run_case forces the Triton family around it), not a public backend
+        return ImplementationType.MINIWORLD if i == "triton" else ImplementationType(i)
 
     # The dims that used to be declared here -- PAIR_D, PAIR_HID, HID_D, BOTH -- are in
     # registry_module.csv now, with a `source` column saying where each number came from. They are
@@ -463,6 +464,21 @@ def cases() -> list[Case]:
 def run_case(case: Case, length: int, dim_index: int, *, train: bool, p_drop: float = 0.0,
              impl: str = "miniworld", dtype: torch.dtype = BF16,
              compute_dtype: torch.dtype | None = None) -> int:
+    """:func:`_run_case`, with the engine's Triton family forced for the ``triton`` impl: it is not a public backend any more, so the unit
+    builds a ``miniworld`` module under ``settings.force_triton_kernels`` (the A/B switch), as the explicit ``triton`` request did."""
+    from miniworld_engine import settings
+
+    previous = settings.configure(force_triton_kernels=True) if impl == "triton" else None
+    try:
+        return _run_case(case, length, dim_index, train=train, p_drop=p_drop, impl=impl, dtype=dtype, compute_dtype=compute_dtype)
+    finally:
+        if previous is not None:
+            settings.configure(**dataclasses.asdict(previous))
+
+
+def _run_case(case: Case, length: int, dim_index: int, *, train: bool, p_drop: float = 0.0,
+              impl: str = "miniworld", dtype: torch.dtype = BF16,
+              compute_dtype: torch.dtype | None = None) -> int:
     """Forward (and backward) once, so every kernel the module dispatches to fires. Returns 1 on
     success, 0 if the module could not run this shape (an unsupported width is data, not failure).
 

miniworld_engine-2.3.10/src/miniworld_engine/autotune/configs/default/augmented_attention_bwd_reduce_window_triton.csv

@@ -0,0 +1,7 @@
+BLOCK_D,num_warps,num_stages
+32,4,2
+64,4,2
+16,4,2
+32,2,2
+32,8,2
+32,4,1

miniworld_engine-2.3.10/src/miniworld_engine/autotune/configs/default/augmented_attention_bwd_window_triton.csv

@@ -0,0 +1,9 @@
+BLOCK_N,num_warps,num_stages
+128,4,2
+64,4,2
+128,8,2
+64,2,2
+32,4,2
+128,4,1
+64,4,3
+128,2,2

miniworld_engine-2.3.10/src/miniworld_engine/autotune/configs/default/augmented_attention_fwd_window_triton.csv

@@ -0,0 +1,9 @@
+BLOCK_N,num_warps,num_stages
+128,4,2
+64,4,2
+128,8,2
+64,2,2
+32,4,2
+128,4,1
+64,4,3
+128,2,2

miniworld_engine-2.3.10/src/miniworld_engine/autotune/configs/default/outer_product_mean_bwd_epilogue_ln_dw_triton.csv

@@ -0,0 +1,33 @@
+BM,BN,BK,SPLIT,num_warps,num_stages
+64,64,64,64,4,2
+128,64,64,64,4,2
+64,128,64,64,4,2
+64,64,128,64,4,2
+64,64,64,128,4,2
+64,64,64,64,8,2
+64,64,64,64,4,3
+32,64,64,64,4,2
+64,32,64,64,4,2
+64,64,32,64,4,2
+64,64,64,32,4,2
+64,64,64,64,2,2
+64,64,64,64,4,1
+16,64,64,64,4,2
+64,256,64,64,4,2
+64,64,16,64,4,2
+64,64,64,16,4,2
+64,64,64,64,1,2
+64,64,64,64,4,4
+64,64,64,8,4,2
+64,64,64,64,4,5
+64,64,64,4,4,2
+64,64,64,64,4,6
+64,64,64,2,4,2
+64,64,64,1,4,2
+64,64,64,64,8,3
+64,64,64,128,4,3
+64,64,64,128,8,2
+64,64,64,128,8,3
+64,64,128,64,4,3
+64,64,128,64,8,2
+64,64,128,64,8,3

miniworld_engine-2.3.10/src/miniworld_engine/autotune/configs/default/outer_product_mean_bwd_epilogue_ln_dx_triton.csv

@@ -0,0 +1,12 @@
+BJ,PPS,num_warps,num_stages
+8,2,4,1
+4,2,4,1
+16,2,4,1
+32,2,4,1
+8,1,4,1
+8,4,4,1
+8,8,4,1
+8,2,8,1
+8,2,2,1
+16,2,8,1
+8,2,4,2

miniworld_engine-2.3.10/src/miniworld_engine/autotune/configs/default/outer_product_mean_epilogue_ln_triton.csv

@@ -0,0 +1,11 @@
+BJ,num_warps,num_stages
+16,4,1
+8,4,1
+32,4,1
+64,4,1
+4,4,1
+16,2,1
+16,8,1
+32,8,1
+8,2,1
+16,4,2

miniworld_engine-2.3.10/src/miniworld_engine/autotune/configs/grid/augmented_attention_bwd_reduce_window_triton.csv

@@ -0,0 +1,4 @@
+axis,values
+BLOCK_D,16 32 64
+num_warps,1 2 4 8
+num_stages,1 2 3 4

miniworld_engine-2.3.10/src/miniworld_engine/autotune/configs/grid/augmented_attention_bwd_window_triton.csv

@@ -0,0 +1,4 @@
+axis,values
+BLOCK_N,16 32 64 128 256
+num_warps,1 2 4 8 16
+num_stages,1 2 3 4

miniworld_engine-2.3.10/src/miniworld_engine/autotune/configs/grid/augmented_attention_fwd_window_triton.csv

@@ -0,0 +1,4 @@
+axis,values
+BLOCK_N,16 32 64 128 256
+num_warps,1 2 4 8 16
+num_stages,1 2 3 4

miniworld_engine-2.3.10/src/miniworld_engine/autotune/configs/grid/outer_product_mean_bwd_epilogue_ln_dw_triton.csv

@@ -0,0 +1,7 @@
+axis,values
+BM,16 32 64 128
+BN,32 64 128 256
+BK,16 32 64 128
+SPLIT,1 2 4 8 16 32 64 128
+num_warps,1 2 4 8
+num_stages,1 2 3 4 5 6

miniworld_engine-2.3.10/src/miniworld_engine/autotune/configs/grid/outer_product_mean_bwd_epilogue_ln_dx_triton.csv

@@ -0,0 +1,5 @@
+axis,values
+BJ,4 8 16 32
+PPS,1 2 4 8
+num_warps,2 4 8
+num_stages,1 2

miniworld_engine-2.3.10/src/miniworld_engine/autotune/configs/grid/outer_product_mean_epilogue_ln_triton.csv

@@ -0,0 +1,4 @@
+axis,values
+BJ,4 8 16 32 64
+num_warps,2 4 8
+num_stages,1 2

{miniworld_engine-2.3.9 → miniworld_engine-2.3.10}/src/miniworld_engine/autotune/derive.py

@@ -562,6 +562,11 @@ def derive_all(rows: list[ModuleRow] | None = None, *, on_unit=None,
     require_environment()
     if arch:
         target_arch(arch)
+    # The derivation asks which kernels the modules launch, not whether this card is allowed to launch them: route as if the
+    # card were recorded (a patched arch above carries a real card's name, which no manifest knows).
+    from miniworld_engine.modules import dispatch
+
+    dispatch.assume_every_card_recorded()
     install_no_calibration()
     install_module_apply()
     install_native_recorders()

{miniworld_engine-2.3.9 → miniworld_engine-2.3.10}/src/miniworld_engine/autotune/manifests/NVIDIA H100 80GB HBM3 (sm90).csv

@@ -1,9 +1,9 @@
 kernel,backend,family,file,status,detail,dtype
-#provenance,2.3.8,7bd3ec119a29,clean,2026-10-22,v2.3.7-11-g7bd3ec11,
-adaln_bwd_dx_dlnw_triton,triton,adaln,miniworld_engine/kernels/adaln/triton/training.py,ok,rel DCond=2.79e-03 DLnW=2.04e-03 (band 1e-02 declared),bf16
-adaln_bwd_pre_dx_triton,triton,adaln,miniworld_engine/kernels/adaln/triton/training.py,ok,rel D=1.51e-03 DX=3.13e-03 (band 1e-02 declared),bf16
-adaln_epilogue_saveact_triton,triton,adaln,miniworld_engine/kernels/adaln/triton/training.py,ok,rel Y=2.33e-03 Mean=2.23e-08 Rstd=1.86e-07 Gate=1.95e-03 (band 1e-02 declared),bf16
-adaln_epilogue_triton,triton,adaln,miniworld_engine/kernels/adaln/triton/inference.py,ok,rel out=2.48e-03 (band 1e-02 declared),bf16
+#provenance,2.3.9,d6d1eb084d32,clean,2026-10-24,v2.3.8-9-gd6d1eb08,
+adaln_bwd_dx_dlnw_triton,triton,adaln,miniworld_engine/kernels/adaln/triton/training.py,ok,rel DCond=2.15e-03 DLnW=2.62e-03 (band 1e-02 declared),bf16
+adaln_bwd_pre_dx_triton,triton,adaln,miniworld_engine/kernels/adaln/triton/training.py,ok,rel D=1.49e-03 DX=3.13e-03 (band 1e-02 declared),bf16
+adaln_epilogue_saveact_triton,triton,adaln,miniworld_engine/kernels/adaln/triton/training.py,ok,rel Y=2.65e-03 Mean=2.23e-08 Rstd=1.86e-07 Gate=1.96e-03 (band 1e-02 declared),bf16
+adaln_epilogue_triton,triton,adaln,miniworld_engine/kernels/adaln/triton/inference.py,ok,rel out=2.69e-03 (band 1e-02 declared),bf16
 adaln_fwd_gate_triton,triton,adaln,miniworld_engine/kernels/adaln/triton/training.py,ok,rel Y=2.14e-03 Gate=1.95e-03 (band 8e-03 declared),bf16
 adaln_fwd_triton,triton,adaln,miniworld_engine/kernels/adaln/triton/inference.py,ok,rel out=3.32e-03 (band 1e-02 declared),bf16
 adaln_gemm_gate_triton,triton,adaln,miniworld_engine/kernels/adaln/triton/inference.py,ok,rel out=2.14e-03 (band 8e-03 declared),bf16
@@ -14,15 +14,15 @@ augmented_attention_fwd_triton,triton,augmented_attention,miniworld_engine/kerne
 bias_only_attention_bwd_pre_triton,triton,bias_only_attention,miniworld_engine/kernels/bias_only_attention/triton/main.py,ok,rel out=6.62e-08 (band 3e-07 declared),bf16
 bias_only_attention_bwd_triton,triton,bias_only_attention,miniworld_engine/kernels/bias_only_attention/triton/main.py,ok,rel dv=7.22e-03 dbias=5.84e-03 (band 5e-02 declared),bf16
 bias_only_attention_fwd_triton,triton,bias_only_attention,miniworld_engine/kernels/bias_only_attention/triton/main.py,ok,rel out=8.88e-03 m=8.11e-04 (band 4e-02 declared),bf16
-cond_transition_bwd_gemm_swiglu_triton,triton,conditioned_transition,miniworld_engine/kernels/conditioned_transition/triton/training.py,ok,rel out=2.98e-03 (band 1e-02 declared),bf16
-cond_transition_bwd_swiglu_flat_triton,triton,conditioned_transition,miniworld_engine/kernels/conditioned_transition/triton/training.py,ok,rel out=2.37e-03 (band 8e-03 declared),bf16
-cond_transition_expand_swiglu_saveact_triton,triton,conditioned_transition,miniworld_engine/kernels/conditioned_transition/triton/fwd_saveact.py,ok,rel H=1.97e-03 AB=2.17e-03 (band 1e-02 declared),bf16
+cond_transition_bwd_gemm_swiglu_triton,triton,conditioned_transition,miniworld_engine/kernels/conditioned_transition/triton/training.py,ok,rel out=2.49e-03 (band 1e-02 declared),bf16
+cond_transition_bwd_swiglu_flat_triton,triton,conditioned_transition,miniworld_engine/kernels/conditioned_transition/triton/training.py,ok,rel out=2.69e-03 (band 8e-03 declared),bf16
+cond_transition_expand_swiglu_saveact_triton,triton,conditioned_transition,miniworld_engine/kernels/conditioned_transition/triton/fwd_saveact.py,ok,rel H=1.97e-03 AB=2.23e-03 (band 1e-02 declared),bf16
 cond_transition_expand_swiglu_triton,triton,conditioned_transition,miniworld_engine/kernels/conditioned_transition/triton/composed.py,ok,rel out=1.97e-03 (band 1e-02 declared),bf16
-cond_transition_fwd_b2b_saveact_triton,triton,conditioned_transition,miniworld_engine/kernels/conditioned_transition/triton/training.py,ok,rel Y=1.77e-03 AB=2.17e-03 H=1.97e-03 Out=1.78e-03 Scale=2.06e-03 (band 3e-03 declared),bf16
-cond_transition_fwd_b2b_triton,triton,conditioned_transition,miniworld_engine/kernels/conditioned_transition/triton/inference.py,ok,rel out=3.23e-03 (band 1e-02 declared),bf16
-cond_transition_squeeze_gate_saveact_triton,triton,conditioned_transition,miniworld_engine/kernels/conditioned_transition/triton/fwd_saveact.py,ok,rel Y=3.22e-03 Out=2.96e-03 Scale=2.06e-03 (band 1e-02 declared),bf16
-cond_transition_squeeze_gate_triton,triton,conditioned_transition,miniworld_engine/kernels/conditioned_transition/triton/composed.py,ok,rel out=3.22e-03 (band 8e-03 declared),bf16
-cond_transition_swiglu_triton,triton,conditioned_transition,miniworld_engine/kernels/conditioned_transition/triton/training.py,ok,rel out=3.98e-03 (band 1e-02 declared),bf16
+cond_transition_fwd_b2b_saveact_triton,triton,conditioned_transition,miniworld_engine/kernels/conditioned_transition/triton/training.py,ok,rel Y=1.88e-03 AB=2.23e-03 H=1.97e-03 Out=1.88e-03 Scale=2.06e-03 (band 3e-03 declared),bf16
+cond_transition_fwd_b2b_triton,triton,conditioned_transition,miniworld_engine/kernels/conditioned_transition/triton/inference.py,ok,rel out=3.42e-03 (band 1e-02 declared),bf16
+cond_transition_squeeze_gate_saveact_triton,triton,conditioned_transition,miniworld_engine/kernels/conditioned_transition/triton/fwd_saveact.py,ok,rel Y=3.43e-03 Out=3.15e-03 Scale=2.06e-03 (band 1e-02 declared),bf16
+cond_transition_squeeze_gate_triton,triton,conditioned_transition,miniworld_engine/kernels/conditioned_transition/triton/composed.py,ok,rel out=3.43e-03 (band 8e-03 declared),bf16
+cond_transition_swiglu_triton,triton,conditioned_transition,miniworld_engine/kernels/conditioned_transition/triton/training.py,ok,rel out=5.71e-03 (band 1e-02 declared),bf16
 gated_projection_bwd_dx_triton,triton,bias_only_attention,miniworld_engine/kernels/bias_only_attention/triton/gate_out.py,ok,rel d_out_r=2.66e-03 d_gate=2.40e-03 gated=2.06e-03 (band 1e-02 declared),bf16
 gated_projection_bwd_gate_dropres_triton,triton,trimul_inproj,miniworld_engine/kernels/trimul_inproj/triton/gate_elem.py,ok,rel d_proj=3.32e-03 d_glogit=2.39e-03 d_proj_drop=2.17e-03 d_glogit_drop=1.79e-03 d_proj_preact=3.57e-03 d_glogit_preact=2.40e-03 (band 1e-02 declared),bf16
 gated_projection_bwd_gate_flat_triton,triton,gated_projection,miniworld_engine/kernels/gated_projection/triton/main.py,ok,rel dg=2.69e-03 do=2.22e-03 (band 1e-02 declared),bf16
@@ -32,7 +32,7 @@ gated_projection_gate_dropres_triton,triton,trimul_inproj,miniworld_engine/kerne
 gated_projection_gate_flat_triton,triton,gated_projection,miniworld_engine/kernels/gated_projection/triton/main.py,ok,rel out=2.15e-03 (band 8e-03 declared),bf16
 gated_projection_gate_gemm_triton,triton,bias_only_attention,miniworld_engine/kernels/bias_only_attention/triton/gate_out.py,ok,rel out=3.79e-03 (band 1e-02 declared),bf16
 gated_projection_gate_triton,triton,gated_projection,miniworld_engine/kernels/gated_projection/triton/main.py,ok,rel out=2.15e-03 (band 8e-03 declared),bf16
-layernorm_bwd_atomic_strided_triton,triton,layernorm_linear,miniworld_engine/kernels/layernorm_linear/triton/te_style.py,ok,rel dx=2.60e-03 dgamma=2.99e-07 dbeta=1.67e-07 (band 1e-02 declared),bf16
+layernorm_bwd_atomic_strided_triton,triton,layernorm_linear,miniworld_engine/kernels/layernorm_linear/triton/te_style.py,ok,rel dx=2.60e-03 dgamma=3.78e-07 dbeta=3.34e-07 (band 1e-02 declared),bf16
 layernorm_bwd_atomic_triton,triton,layernorm,miniworld_engine/kernels/layernorm/triton/main.py,ok,rel dx=2.03e-03 dw=2.26e-03 db=2.41e-03 (band 1e-02 declared),bf16
 layernorm_bwd_foldstats_triton,triton,transition,miniworld_engine/kernels/transition/triton/fused.py,ok,rel dx=2.61e-03 dgamma=2.40e-07 dbeta=8.45e-08 (band 9e-03 declared),bf16
 layernorm_bwd_reduce_cuda,cuda,layernorm,miniworld_engine/kernels/layernorm/cuda/layer_norm_cuda_kernel.cu,ok,rel dw=2.26e-03 db=2.41e-03 (band 9e-03 declared),bf16
@@ -45,11 +45,11 @@ layernorm_fwd_recompute_noaffine_triton,triton,layernorm_linear,miniworld_engine
 layernorm_fwd_recompute_triton,triton,layernorm_linear,miniworld_engine/kernels/layernorm_linear/triton/recompute.py,ok,rel out=2.11e-03 (band 1e-02 declared),bf16
 layernorm_fwd_saveact_strided_triton,triton,layernorm_linear,miniworld_engine/kernels/layernorm_linear/triton/te_style.py,ok,rel x_normed=2.10e-03 mean=4.71e-08 rstd=1.82e-07 (band 1e-02 declared),bf16
 layernorm_fwd_saveact_triton,triton,layernorm,miniworld_engine/kernels/layernorm/triton/main.py,ok,rel out=2.11e-03 (band 1e-02 declared),bf16
-layernorm_fwd_strided_triton,triton,adaln,miniworld_engine/kernels/adaln/triton/ln_strided.py,ok,rel out=2.64e-03 (band 9e-03 declared),bf16
+layernorm_fwd_strided_triton,triton,adaln,miniworld_engine/kernels/adaln/triton/ln_strided.py,ok,rel out=2.56e-03 (band 9e-03 declared),bf16
 layernorm_linear_bwd_fp32_triton,triton,layernorm_linear,miniworld_engine/kernels/layernorm_linear/triton/pair_bias.py,ok,rel dx=2.30e-03 dln_weight=1.74e-03 dproj_weight=2.01e-03 (band 1e-02 declared),bf16
 layernorm_linear_fwd_fp32_triton,triton,layernorm_linear,miniworld_engine/kernels/layernorm_linear/triton/pair_bias.py,ok,rel out=2.37e-03 mean=2.08e-08 rstd=1.74e-07 (band 9e-03 declared),bf16
 layernorm_linear_fwd_triton,triton,layernorm_linear,miniworld_engine/kernels/layernorm_linear/triton/fused.py,ok,rel out=3.51e-03 (band 2e-02 declared),bf16
-layernorm_stats_triton,triton,layernorm_linear,miniworld_engine/kernels/layernorm_linear/triton/stats.py,ok,rel rstd=8.69e-08 c1=1.63e-07 (band 6e-07 declared),bf16
+layernorm_stats_triton,triton,layernorm_linear,miniworld_engine/kernels/layernorm_linear/triton/stats.py,ok,rel rstd=8.69e-08 c1=8.17e-08 (band 6e-07 declared),bf16
 outer_product_mean_bwd_epilogue_dw_triton,triton,outer_product_mean,miniworld_engine/kernels/outer_product_mean/triton/main.py,ok,rel dw_out=4.55e-03 (band 1e-02 declared),bf16
 outer_product_mean_bwd_epilogue_dx_triton,triton,outer_product_mean,miniworld_engine/kernels/outer_product_mean/triton/main.py,ok,rel dmsa=5.84e-03 db_out=2.05e-03 dresidual=0.00e+00 (band 1e-02 declared),bf16
 outer_product_mean_bwd_layernorm_gemm_triton,triton,outer_product_mean,miniworld_engine/kernels/outer_product_mean/triton/main.py,ok,rel dmsa=5.84e-03 dln_weight=4.26e-03 dln_bias=3.54e-03 dw_left=4.97e-03 dw_right=6.01e-03 (band 1e-02 declared),bf16
@@ -99,7 +99,7 @@ transition_fwd_b2b_sm90_cuda,cuda,transition,miniworld_engine/kernels/transition
 transition_expand_gate_sm90_cuda,cuda,transition,miniworld_engine/kernels/transition/cuda/transition_expand_gate_kernel.cu,ok,rel out=1.94e-03 (band 8e-03 declared),bf16
 transition_bwd_gate_sm90_cuda,cuda,transition,miniworld_engine/kernels/transition/cuda/transition_gatebwd_kernel.cu,ok,rel h=1.94e-03 dAB=2.97e-03 xn=5.27e-04 (band 1e-02 declared),bf16
 trimul_output_f567_train_triton,triton,trimul_inproj,miniworld_engine/kernels/trimul_inproj/triton/output_fused.py,ok,rel y=4.27e-03 proj=2.95e-03 gate=2.72e-03 (band 1e-02 declared),bf16
-trimul_input_ln_residual_bwd_triton,triton,trimul_inproj,miniworld_engine/kernels/trimul_inproj/triton/backward_fused.py,ok,rel dx=2.29e-03 dw=4.02e-07 db=1.63e-07 (band 1e-02 declared),bf16
+trimul_input_ln_residual_bwd_triton,triton,trimul_inproj,miniworld_engine/kernels/trimul_inproj/triton/backward_fused.py,ok,rel dx=2.29e-03 dw=2.41e-07 db=8.17e-08 (band 1e-02 declared),bf16
 trimul_input_dual_bwd_triton,triton,trimul_inproj,miniworld_engine/kernels/trimul_inproj/triton/backward_fused.py,ok,rel dx=2.34e-03 (band 1e-02 declared),bf16
 transition_squeeze_residual_triton,triton,transition,miniworld_engine/kernels/transition/triton/residual.py,ok,rel out=2.14e-03 (band 1e-02 declared),bf16
 mpnn_edge_tail_fwd_gemm_gather_saveact_triton,triton,mpnn_edge_tail,miniworld_engine/kernels/mpnn_edge_tail/triton/compute.py,ok,rel out=6.01e-03 (band 5e-02),bf16
@@ -115,7 +115,7 @@ mpnn_edge_tail_bwd_dx_recompute_triton,triton,mpnn_edge_tail,miniworld_engine/ke
 mpnn_node_message_fwd_gemm_triton,triton,mpnn_node_message,miniworld_engine/kernels/mpnn_node_message/triton/main.py,ok,rel out=1.49e-03 (band 5e-02),bf16
 mpnn_node_message_bwd_recompute_triton,triton,mpnn_node_message,miniworld_engine/kernels/mpnn_node_message/triton/main.py,failed,WRONG NUMBERS: rel edge_states=1.10e-01 query_projection=1.02e-02 neighbor_projection=3.51e-02 edge_weight=2.20e-02 hidden_weight=6.86e-03 hidden_bias=3.82e-03 (band 5e-02),bf16
 mpnn_node_message_bwd_dx_triton,triton,mpnn_node_message,miniworld_engine/kernels/mpnn_node_message/triton/main.py,failed,WRONG NUMBERS: rel edge_states=1.10e-01 query_projection=1.02e-02 neighbor_projection=3.51e-02 edge_weight=2.20e-02 hidden_weight=6.86e-03 hidden_bias=3.82e-03 (band 5e-02),bf16
-mpnn_relative_position_bwd_reduce_triton,triton,mpnn_relative_position,miniworld_engine/kernels/mpnn_relative_position/triton/main.py,ok,rel grad_table=1.15e-07 grad_bias=7.24e-08 (band 5e-02),bf16
+mpnn_relative_position_bwd_reduce_triton,triton,mpnn_relative_position,miniworld_engine/kernels/mpnn_relative_position/triton/main.py,ok,rel grad_table=5.77e-08 grad_bias=7.24e-08 (band 5e-02),bf16
 mpnn_message_fwd_gemm_triton,triton,mpnn_message,miniworld_engine/kernels/mpnn_message/triton/main.py,ok,rel out=1.07e-03 (band 5e-02),bf16
 mpnn_message_fwd_gelu_reduce_triton,triton,mpnn_message,miniworld_engine/kernels/mpnn_message/triton/main.py,ok,rel out=1.07e-03 (band 5e-02),bf16
 mpnn_message_bwd_dx_triton,triton,mpnn_message,miniworld_engine/kernels/mpnn_message/triton/main.py,ok,rel preactivation=1.01e-02 weight=4.23e-03 bias=3.97e-03 (band 5e-02),bf16
@@ -127,22 +127,22 @@ mpnn_edge_dropout_bwd_packed_triton,triton,mpnn_edge_dropout,miniworld_engine/ke
 trimul_fwd_sm90_cuda,cuda,trimul_inproj,miniworld_engine/kernels/trimul_inproj/cuda/h100_inference.py,ok,rel h128_dir0=4.58e-03 h128_dir1=4.64e-03 h128_dir2=4.63e-03 h256_dir0=5.06e-03 h256_dir1=4.88e-03 h256_dir2=5.34e-03 (band 2e-02 declared),bf16
 transition_fwd_residual_sm90_cuda,cuda,transition,miniworld_engine/kernels/transition/cuda/transition_fused_fwd_sm90a_kernel.cu,ok,rel n4_y=3.57e-03 n4_xn=3.10e-03 n4_rstd=8.69e-08 n4_c1=6.69e-08 n2_y=3.33e-03 n2_xn=3.10e-03 n2_rstd=8.69e-08 n2_c1=6.69e-08 (band 1e-02 declared),bf16
 transition_bwd_residual_sm90_cuda,cuda,transition,miniworld_engine/kernels/transition/cuda/transition_fused_bwd_sm90a_kernel.cu,ok,rel n4_dx=5.85e-03 n4_dgamma=4.31e-03 n4_dbeta=3.03e-03 n4_dWa=3.97e-03 n4_dWb=4.61e-03 n4_dWs=4.17e-03 n2_dx=6.01e-03 n2_dgamma=4.39e-03 n2_dbeta=3.81e-03 n2_dWa=3.92e-03 n2_dWb=4.27e-03 n2_dWs=3.68e-03 (band 2e-02 declared),bf16
-swa_dit_inproj_fwd_triton,triton,swa_dit,miniworld_engine/kernels/swa_dit/triton/forward.py,ok,rel out=5.90e-03 (band 5e-02 declared),bf16
-swa_dit_softmax_fwd_triton,triton,swa_dit,miniworld_engine/kernels/swa_dit/triton/forward.py,ok,rel out=5.90e-03 (band 5e-02 declared),bf16
-swa_dit_output_swiglu_fwd_triton,triton,swa_dit,miniworld_engine/kernels/swa_dit/triton/forward.py,ok,rel out=5.90e-03 (band 5e-02 declared),bf16
+swa_dit_inproj_fwd_triton,triton,swa_dit,miniworld_engine/kernels/swa_dit/triton/forward.py,ok,rel out=5.94e-03 (band 5e-02 declared),bf16
+swa_dit_softmax_fwd_triton,triton,swa_dit,miniworld_engine/kernels/swa_dit/triton/forward.py,ok,rel out=5.94e-03 (band 5e-02 declared),bf16
+swa_dit_output_swiglu_fwd_triton,triton,swa_dit,miniworld_engine/kernels/swa_dit/triton/forward.py,ok,rel out=5.94e-03 (band 5e-02 declared),bf16
 swa_dit_swiglu_bwd_triton,triton,swa_dit,miniworld_engine/kernels/swa_dit/triton/backward.py,ok,rel dmod=3.40e-03 dw_up=4.71e-03 dw_down=4.70e-03 (band 5e-02 declared),bf16
 swa_dit_swiglu_dw_triton,triton,swa_dit,miniworld_engine/kernels/swa_dit/triton/backward.py,ok,rel dw_up=4.71e-03 dw_down=4.70e-03 (band 5e-02 declared),bf16
-swa_dit_output_bwd_triton,triton,swa_dit,miniworld_engine/kernels/swa_dit/triton/backward.py,ok,rel dmod=3.40e-03 dw_out=4.32e-03 (band 5e-02 declared),bf16
-swa_dit_softmax_bwd_dq_triton,triton,swa_dit,miniworld_engine/kernels/swa_dit/triton/backward.py,ok,rel dq=6.18e-03 dw_qkv=6.74e-03 (band 5e-02 declared),bf16
-swa_dit_softmax_bwd_dkdv_triton,triton,swa_dit,miniworld_engine/kernels/swa_dit/triton/backward.py,ok,rel dq=6.18e-03 dw_qkv=6.74e-03 (band 5e-02 declared),bf16
-swa_dit_inproj_bwd_triton,triton,swa_dit,miniworld_engine/kernels/swa_dit/triton/backward.py,ok,rel dq=6.18e-03 dmod=3.40e-03 dw_qkv=6.74e-03 dw_gate=4.86e-03 (band 5e-02 declared),bf16
+swa_dit_output_bwd_triton,triton,swa_dit,miniworld_engine/kernels/swa_dit/triton/backward.py,ok,rel dmod=3.40e-03 dw_out=4.66e-03 (band 5e-02 declared),bf16
+swa_dit_softmax_bwd_dq_triton,triton,swa_dit,miniworld_engine/kernels/swa_dit/triton/backward.py,ok,rel dq=6.18e-03 dw_qkv=5.72e-03 (band 5e-02 declared),bf16
+swa_dit_softmax_bwd_dkdv_triton,triton,swa_dit,miniworld_engine/kernels/swa_dit/triton/backward.py,ok,rel dq=6.18e-03 dw_qkv=5.72e-03 (band 5e-02 declared),bf16
+swa_dit_inproj_bwd_triton,triton,swa_dit,miniworld_engine/kernels/swa_dit/triton/backward.py,ok,rel dq=6.18e-03 dmod=3.40e-03 dw_qkv=5.72e-03 dw_gate=4.68e-03 (band 5e-02 declared),bf16
 swa_dit_inproj_fwd_sm90_cuda,cuda,swa_dit,miniworld_engine/kernels/swa_dit/cuda/swa_qkvg_fwd.cu,ok,rel out=5.90e-03 (band 5e-02 declared),bf16
-swa_dit_output_swiglu_fwd_sm90_cuda,cuda,swa_dit,miniworld_engine/kernels/swa_dit/cuda/swa_ffn_fwd.cu,ok,rel out=5.90e-03 (band 5e-02 declared),bf16
+swa_dit_output_swiglu_fwd_sm90_cuda,cuda,swa_dit,miniworld_engine/kernels/swa_dit/cuda/swa_ffn_fwd.cu,ok,rel out=5.94e-03 (band 5e-02 declared),bf16
 swa_dit_swiglu_bwd_sm90_cuda,cuda,swa_dit,miniworld_engine/kernels/swa_dit/cuda/swa_ffn_bwd.cu,ok,rel dq=6.18e-03 dmod=3.40e-03 dw_up=4.71e-03 dw_down=4.70e-03 (band 5e-02 declared),bf16
 swa_dit_softmax_fwd_sm90_cuda,cuda,swa_dit,miniworld_engine/kernels/swa_dit/cuda/swa_attn.cu,ok,rel out=5.90e-03 (band 5e-02 declared),bf16
 swa_dit_softmax_bwd_sm90_cuda,cuda,swa_dit,miniworld_engine/kernels/swa_dit/cuda/swa_attn.cu,ok,rel dq=5.60e-03 dmod=3.64e-03 dw_qkv=5.72e-03 (band 5e-02 declared),bf16
-swa_dit_output_bwd_sm90_cuda,cuda,swa_dit,miniworld_engine/kernels/swa_dit/cuda/swa_slab.cu,ok,rel dq=6.18e-03 dmod=3.40e-03 dw_out=4.32e-03 dw_gate=4.86e-03 (band 5e-02 declared),bf16
-swa_dit_inproj_bwd_sm90_cuda,cuda,swa_dit,miniworld_engine/kernels/swa_dit/cuda/swa_slab.cu,ok,rel dq=6.18e-03 dmod=3.40e-03 dw_qkv=6.74e-03 dw_gate=4.86e-03 (band 5e-02 declared),bf16
+swa_dit_output_bwd_sm90_cuda,cuda,swa_dit,miniworld_engine/kernels/swa_dit/cuda/swa_slab.cu,ok,rel dq=6.18e-03 dmod=3.40e-03 dw_out=4.66e-03 dw_gate=4.68e-03 (band 5e-02 declared),bf16
+swa_dit_inproj_bwd_sm90_cuda,cuda,swa_dit,miniworld_engine/kernels/swa_dit/cuda/swa_slab.cu,ok,rel dq=6.18e-03 dmod=3.40e-03 dw_qkv=5.72e-03 dw_gate=4.68e-03 (band 5e-02 declared),bf16
 swa_dit_inproj_fwd_rows_sm90_cuda,cuda,swa_dit,miniworld_engine/kernels/swa_dit/cuda/swa_slab.cu,ok,rel out=5.90e-03 (band 5e-02 declared),bf16
 swa_dit_output_fwd_rows_sm90_cuda,cuda,swa_dit,miniworld_engine/kernels/swa_dit/cuda/swa_slab.cu,ok,rel out=5.90e-03 (band 5e-02 declared),bf16
 swa_dit_swiglu_fwd_rows_sm90_cuda,cuda,swa_dit,miniworld_engine/kernels/swa_dit/cuda/swa_slab.cu,ok,rel out=5.90e-03 (band 5e-02 declared),bf16

miniworld_engine-2.3.10/src/miniworld_engine/autotune/records.py

@@ -0,0 +1,193 @@
+"""Which cards the engine holds Triton evidence for -- the one question ``modules.dispatch`` asks.
+
+``implementation="miniworld"`` on a card without a native CUDA path routes to the engine's Triton kernel family only if
+the engine has a RECORD for that exact card; otherwise it warns and runs the PyTorch reference. "Recorded" is not a
+judgement made here, it is a lookup of two facts the repo already keeps:
+
+  * ``autotune/manifests/<card>.csv`` -- written by ``python -m miniworld_engine.autotune.run_all`` on the card. A row with
+    status ``ok`` means the kernel was driven on that card and checked against its reference (the ``#provenance`` row says
+    when and at which commit); ``failed`` / ``untested`` rows and kernels with no row at all are NOT evidence.
+  * ``kernels/registry/registry.csv`` -- what exists. The Triton rows of the kernel families an op launches
+    (``OP_KERNELS`` below) are the denominator, minus the ``developed == no`` rows (A/B alternatives no module launches).
+
+An op is recorded on a card FOR A MODE when every such kernel the mode launches has an ``ok`` row for that card at the dtype in use
+(a kernel whose registry row does not declare the dtype does not count against it). The modes are the two calls a module makes:
+``inference`` (no backward is recorded: forward kernels only) and ``training`` (everything). A kernel is a training kernel when its
+name says it saves for or runs the backward -- a ``bwd``, ``saveact``, ``train`` or ``dw`` word (``kernel_mode``); a missing
+backward kernel therefore blocks training calls only, and inference stays on Triton on a card whose forward kernels are recorded. The card is identified by its exact name and
+capability, ``"NVIDIA RTX A6000 (sm86)"``: a card with the same capability and another name (A40, L40S) is not
+recorded -- its shared memory, bandwidth and clocks differ, and nothing was run on it.
+
+Cheap and import-safe: stdlib only (no torch, no GPU), the registry and each manifest are read once. Tests inject a
+fake manifest directory with :func:`use_manifest_dir`.
+"""
+
+from __future__ import annotations
+
+import contextlib
+import csv
+import fnmatch
+import functools
+from collections.abc import Iterator
+from pathlib import Path
+
+_PKG = Path(__file__).resolve().parent.parent
+_REGISTRY = _PKG / "kernels" / "registry" / "registry.csv"
+_MANIFESTS = _PKG / "autotune" / "manifests"
+_PROVENANCE = "#provenance"
+#: A manifest row without a ``dtype`` cell was written before the column existed: bf16 (see ``autotune.devices``).
+_LEGACY_DTYPE = "bf16"
+
+#: Op -> ``family/kernel`` patterns (``fnmatch``) over the registry's Triton rows: the kernels the op's ``miniworld`` route
+#: launches on a card with no native CUDA path. The op names are the ``modules.dispatch`` ones. Family is the registry's
+#: organisational column, not the launcher, so a family shared by two ops is split by kernel name (``gated_projection``:
+#: the pair/token gate GEMMs belong to TriangleAttention, ``swa_gate_out_fwd`` to the SWA atom attention).
+#: The token attention core's kernels (``augmented_attention/triton/main.py``), spelled out because the same family holds the
+#: block-local window kernels (``*_window_triton``), which belong to the atom ops below and must not be held against the token ones.
+_TOKEN_ATTENTION = (
+    "augmented_attention/augmented_attention_fwd_triton",
+    "augmented_attention/augmented_attention_bwd_pre_triton",
+    "augmented_attention/augmented_attention_bwd_reduce_triton",
+    "augmented_attention/augmented_attention_bwd_split_triton",
+    "augmented_attention/augmented_attention_bwd_atomic_triton",
+)
+#: The block-local window attention (32 queries x 128 / 192 keys per window): the atom DiT block's core.
+_WINDOW_ATTENTION = ("augmented_attention/augmented_attention_*_window_triton",)
+#: The inference pair-bias producer of the portable DiT paths: LayerNorm of the pair and its projection to the heads in one pass.
+_PAIR_BIAS = ("layernorm_linear/layernorm_linear_fwd_triton",)
+
+OP_KERNELS: dict[str, tuple[str, ...]] = {
+    "transition": ("transition/*",),
+    "triangle_multiplication": ("trimul_inproj/*",),
+    "triangle_attention": (
+        "triangle_attention/*",
+        "layernorm_linear/*",
+        "layernorm/*",
+        "gated_projection/gated_projection_*",
+        "bias_only_attention/*",
+    ),
+    "adaptive_layernorm": ("adaln/*",),
+    "conditioned_transition": ("conditioned_transition/*", "adaln/*"),
+    "augmented_attention": (*_TOKEN_ATTENTION, "adaln/*"),
+    "outer_product_mean": ("outer_product_mean/*",),
+    "msa_pair_weighted_averaging": ("pair_weighted_averaging/*",),
+    "rmsnorm": ("rmsnorm/*",),
+    "layernorm": ("layernorm/*",),
+    # the ops.* facade entry points that launch a single Triton unit (kernels/*/whole_op.py)
+    "gated_linear": ("gated_projection/gated_projection_*", "bias_only_attention/*"),
+    "augmented_attention_pair_bias": _TOKEN_ATTENTION,
+    "swiglu_ffn": ("transition/transition_expand_swiglu_triton",),
+    "swa_atom_attention": ("rope/*", "rmsnorm/*", "gated_projection/swa_gate_out_fwd_triton"),
+    "swa_dit": ("swa_dit/*", "rmsnorm/*"),
+    # The attention cores of the portable DiT paths (the rest of those blocks is torch ops and the ops above, each with its own rule):
+    # the AF3 / Protenix / OpenFold3 / ESMFold2 atom block (LocalDiTBlock), and the Boltz-2 / RoseTTAFold3 / Chai-1 token and atom stacks.
+    "local_dit": (*_WINDOW_ATTENTION, *_PAIR_BIAS),
+    "dit_variant": (*_TOKEN_ATTENTION, *_PAIR_BIAS),
+    "dit_variant_atom": (*_WINDOW_ATTENTION, *_PAIR_BIAS),
+}
+
+#: The words of a kernel name that make it a training kernel: it saves activations for the backward (`saveact`), is the backward
+#: (`bwd`, `dw`) or the fused training epilogue (`train`). Every other kernel of an op runs in the forward of an inference call too.
+_TRAINING_WORDS = frozenset({"bwd", "saveact", "train", "dw"})
+
+_manifest_dir: Path | None = None
+
+
+def kernel_mode(kernel: str) -> str:
+    """``"training"`` if ``kernel`` only runs where a backward is recorded, else ``"inference"`` (its forward serves both)."""
+    return "training" if _TRAINING_WORDS & set(kernel.split("_")) else "inference"
+
+
+def card_key(name: str, capability: tuple[int, int]) -> str:
+    """``"NVIDIA RTX A6000 (sm86)"``: the manifest / cache key of a card (``autotune.cache.gpu_key`` from arguments)."""
+    return f"{name} (sm{capability[0]}{capability[1]})".replace("/", "_")
+
+
+def _dir() -> Path:
+    return _manifest_dir if _manifest_dir is not None else _MANIFESTS
+
+
+@contextlib.contextmanager
+def use_manifest_dir(path: str | Path | None) -> Iterator[None]:
+    """Read the manifests from ``path`` instead of the packaged ones (tests). ``None`` restores the packaged directory."""
+    global _manifest_dir
+    previous = _manifest_dir
+    _manifest_dir = Path(path) if path is not None else None
+    _load.cache_clear()
+    try:
+        yield
+    finally:
+        _manifest_dir = previous
+        _load.cache_clear()
+
+
+@functools.cache
+def _registry() -> tuple[dict, ...]:
+    with _REGISTRY.open(newline="") as handle:
+        return tuple(csv.DictReader(handle))
+
+
+@functools.cache
+def _load(card: str) -> dict[tuple[str, str], str]:
+    """``{(kernel, dtype): status}`` for ``card``; empty when the card has no manifest."""
+    path = _dir() / f"{card}.csv"
+    if not path.is_file():
+        return {}
+    with path.open(newline="") as handle:
+        rows = [r for r in csv.DictReader(handle) if r["kernel"] != _PROVENANCE]
+    return {(r["kernel"], (r.get("dtype") or "").strip() or _LEGACY_DTYPE): r["status"] for r in rows}
+
+
+def recorded_cards() -> tuple[str, ...]:
+    """Every card with a manifest in the active directory."""
+    return tuple(sorted(p.stem for p in _dir().glob("*.csv")))
+
+
+def triton_kernels(op: str, dtype: str = "bf16", mode: str = "training") -> tuple[str, ...]:
+    """The Triton kernels ``op``'s ``miniworld`` route launches in ``mode`` (developed rows that declare ``dtype``; ``"inference"``
+    leaves out the training kernels, ``"training"`` is all of them)."""
+    if mode not in ("inference", "training"):
+        raise ValueError(f"mode must be 'inference' or 'training', not {mode!r}")
+    patterns = OP_KERNELS.get(op)
+    if patterns is None:
+        return ()
+    out = []
+    for row in _registry():
+        if row["backend"] != "triton" or row["developed"] != "yes":
+            continue
+        declared = {d.strip() for d in (row.get("dtypes") or "bf16").split("|") if d.strip()}
+        if dtype not in declared:
+            continue
+        if mode == "inference" and kernel_mode(row["kernel"]) == "training":
+            continue
+        if any(fnmatch.fnmatchcase(f"{row['family']}/{row['kernel']}", p) for p in patterns):
+            out.append(row["kernel"])
+    return tuple(out)
+
+
+def missing(op: str, name: str, capability: tuple[int, int], dtype: str = "bf16", mode: str = "training") -> tuple[str, ...]:
+    """The kernels ``op`` launches in ``mode`` with no ``ok`` row for this exact card at ``dtype`` (empty means recorded). An op the
+    table does not know has no kernels to vouch for: its only entry is a note naming the op."""
+    kernels = triton_kernels(op, dtype, mode)
+    if not kernels:
+        return (f"<no Triton kernel of {op!r} declares {dtype}>",)
+    rows = _load(card_key(name, capability))
+    return tuple(k for k in kernels if rows.get((k, dtype)) != "ok")
+
+
+def recorded(op: str, name: str, capability: tuple[int, int], dtype: str = "bf16", mode: str = "training") -> bool:
+    """Does the engine hold evidence that the Triton kernels ``op`` launches in ``mode`` run correctly on exactly this card at ``dtype``?"""
+    return not missing(op, name, capability, dtype, mode)
+
+
+def routing_table(dtype: str = "bf16", mode: str = "training") -> dict[str, dict[str, tuple[str, ...]]]:
+    """``{card: {op: missing kernels}}`` over every recorded card -- the data behind ``docs/dispatch.md``."""
+    return {
+        c: {op: missing(op, c.rsplit(" (sm", 1)[0], _capability_of(c), dtype, mode) for op in OP_KERNELS}
+        for c in recorded_cards()
+    }
+
+
+def _capability_of(card: str) -> tuple[int, int]:
+    digits = card.rsplit("(sm", 1)[1].rstrip(")")
+    return int(digits[:-1]), int(digits[-1])