miniworld-engine 2.4.2__tar.gz → 2.4.3__tar.gz

{miniworld_engine-2.4.2 → miniworld_engine-2.4.3}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: miniworld-engine
-Version: 2.4.2
+Version: 2.4.3
 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

{miniworld_engine-2.4.2 → miniworld_engine-2.4.3}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "miniworld-engine"
-version = "2.4.2"
+version = "2.4.3"
 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-2.4.2 → miniworld_engine-2.4.3}/src/miniworld_engine/autotune/manifests/NVIDIA H100 80GB HBM3 (sm90).csv

@@ -1,5 +1,5 @@
 kernel,backend,family,file,status,detail,dtype
-#provenance,2.4.2,f35260f31a37,clean,2026-11-02,v2.4.0-96-gf35260f31,
+#provenance,2.4.3,2142f3921a07,clean,2026-11-02,v2.4.2-9-g2142f3921,
 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
@@ -73,7 +73,7 @@ rmsnorm_adamod_fwd_triton,triton,rmsnorm_adamod,miniworld_engine/kernels/rmsnorm
 rmsnorm_bwd_triton,triton,rmsnorm,miniworld_engine/kernels/rmsnorm/triton/main.py,ok,rel dx_aff=2.03e-03 dweight_aff=2.73e-03 dx_plain=2.54e-03 (band 4e-03 declared),bf16
 rmsnorm_fwd_triton,triton,rmsnorm,miniworld_engine/kernels/rmsnorm/triton/main.py,ok,rel y_aff=2.37e-03 y_plain=3.80e-03 (band 4e-03 declared),bf16
 rope_fwd_triton,triton,rope,miniworld_engine/kernels/rope/triton/main.py,ok,rel y=5.14e-04 dx=7.23e-04 (band 3e-03 declared),bf16
-transition_bwd_cuda,cuda,transition,miniworld_engine/kernels/transition/cuda/transition_cuda_kernel.cu,failed,RuntimeError: Error building extension 'transition_cuda_ext_v2': [1/2] c++ -MMD -MF transition_cuda.o.d -DTORCH_EXTENSION_NAME=transit | c++: error: unrecognized command line option ‘-std=c++20’; did you mean ‘-std=c++2a’?,bf16
+transition_bwd_cuda,cuda,transition,miniworld_engine/kernels/transition/cuda/transition_cuda_kernel.cu,failed,RuntimeError: Error building extension 'transition_cuda_ext_v2': [1/3] c++ -MMD -MF transition_cuda.o.d -DTORCH_EXTENSION_NAME=transit | c++: error: unrecognized command line option ‘-std=c++20’; did you mean ‘-std=c++2a’?,bf16
 transition_bwd_swiglu_recompute_triton,triton,transition,miniworld_engine/kernels/transition/triton/fused.py,ok,rel h=2.18e-03 dAB=2.12e-03 xn=7.02e-04 (band 1e-02 declared),bf16
 transition_bwd_squeeze_swiglu_triton,triton,transition,miniworld_engine/kernels/transition/triton/training.py,ok,rel h=2.15e-03 dA=2.49e-03 dB=2.44e-03 dA_saved=2.54e-03 dB_saved=2.46e-03 (band 1e-02 declared),bf16
 transition_cast_cuda,cuda,transition,miniworld_engine/kernels/transition/cuda/transition_cuda_kernel.cu,failed,ImportError: <path>,bf16
@@ -112,16 +112,16 @@ transition_squeeze_residual_triton,triton,transition,miniworld_engine/kernels/tr
 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
 mpnn_edge_tail_fwd_gemm_b2b_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
 mpnn_edge_tail_fwd_gemm_layernorm_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
-mpnn_edge_tail_bwd_layernorm_saveact_triton,triton,mpnn_edge_tail,miniworld_engine/kernels/mpnn_edge_tail/triton/compute.py,failed,WRONG NUMBERS: rel edge_states=5.99e-02 query_projection=1.95e-02 neighbor_projection=1.72e-02 edge_weight=1.32e-02 hidden_weight=1.66e-02 hidden_bias=1.83e-02 output_weight=3.93e-03 output_bias=1.79e-03 norm_weight=2.17e-03 norm_bias=2.93e-03 (band 5e-02),bf16
-mpnn_edge_tail_bwd_dx_saveact_triton,triton,mpnn_edge_tail,miniworld_engine/kernels/mpnn_edge_tail/triton/compute.py,failed,WRONG NUMBERS: rel edge_states=5.99e-02 query_projection=1.95e-02 neighbor_projection=1.72e-02 edge_weight=1.32e-02 hidden_weight=1.66e-02 hidden_bias=1.83e-02 output_weight=3.93e-03 output_bias=1.79e-03 norm_weight=2.17e-03 norm_bias=2.93e-03 (band 5e-02),bf16
-mpnn_edge_tail_bwd_dx_gather_saveact_triton,triton,mpnn_edge_tail,miniworld_engine/kernels/mpnn_edge_tail/triton/compute.py,failed,WRONG NUMBERS: rel edge_states=5.99e-02 query_projection=1.95e-02 neighbor_projection=1.72e-02 edge_weight=1.32e-02 hidden_weight=1.66e-02 hidden_bias=1.83e-02 output_weight=3.93e-03 output_bias=1.79e-03 norm_weight=2.17e-03 norm_bias=2.93e-03 (band 5e-02),bf16
+mpnn_edge_tail_bwd_layernorm_saveact_triton,triton,mpnn_edge_tail,miniworld_engine/kernels/mpnn_edge_tail/triton/compute.py,ok,rel edge_states=5.99e-02 query_projection=1.95e-02 neighbor_projection=1.72e-02 edge_weight=1.32e-02 hidden_weight=1.66e-02 hidden_bias=1.83e-02 output_weight=3.93e-03 output_bias=1.79e-03 norm_weight=2.17e-03 norm_bias=2.93e-03 (band 1e-01 declared),bf16
+mpnn_edge_tail_bwd_dx_saveact_triton,triton,mpnn_edge_tail,miniworld_engine/kernels/mpnn_edge_tail/triton/compute.py,ok,rel edge_states=5.99e-02 query_projection=1.95e-02 neighbor_projection=1.72e-02 edge_weight=1.32e-02 hidden_weight=1.66e-02 hidden_bias=1.83e-02 output_weight=3.93e-03 output_bias=1.79e-03 norm_weight=2.17e-03 norm_bias=2.93e-03 (band 1e-01 declared),bf16
+mpnn_edge_tail_bwd_dx_gather_saveact_triton,triton,mpnn_edge_tail,miniworld_engine/kernels/mpnn_edge_tail/triton/compute.py,ok,rel edge_states=5.99e-02 query_projection=1.95e-02 neighbor_projection=1.72e-02 edge_weight=1.32e-02 hidden_weight=1.66e-02 hidden_bias=1.83e-02 output_weight=3.93e-03 output_bias=1.79e-03 norm_weight=2.17e-03 norm_bias=2.93e-03 (band 1e-01 declared),bf16
 mpnn_edge_tail_fwd_gemm_recompute_triton,triton,mpnn_edge_tail,miniworld_engine/kernels/mpnn_edge_tail/triton/main.py,ok,rel out=5.59e-03 (band 5e-02),bf16
 mpnn_edge_tail_fwd_gemm_layernorm_recompute_triton,triton,mpnn_edge_tail,miniworld_engine/kernels/mpnn_edge_tail/triton/main.py,ok,rel out=5.59e-03 (band 5e-02),bf16
-mpnn_edge_tail_bwd_recompute_triton,triton,mpnn_edge_tail,miniworld_engine/kernels/mpnn_edge_tail/triton/main.py,failed,WRONG NUMBERS: rel edge_states=1.27e-01 query_projection=3.43e-02 neighbor_projection=3.28e-02 edge_weight=2.17e-02 hidden_weight=2.10e-02 hidden_bias=2.66e-02 output_weight=5.13e-03 output_bias=1.79e-03 norm_weight=4.02e-03 norm_bias=2.93e-03 (band 5e-02),bf16
-mpnn_edge_tail_bwd_dx_recompute_triton,triton,mpnn_edge_tail,miniworld_engine/kernels/mpnn_edge_tail/triton/main.py,failed,WRONG NUMBERS: rel edge_states=1.27e-01 query_projection=3.43e-02 neighbor_projection=3.28e-02 edge_weight=2.17e-02 hidden_weight=2.10e-02 hidden_bias=2.66e-02 output_weight=5.13e-03 output_bias=1.79e-03 norm_weight=4.02e-03 norm_bias=2.93e-03 (band 5e-02),bf16
+mpnn_edge_tail_bwd_recompute_triton,triton,mpnn_edge_tail,miniworld_engine/kernels/mpnn_edge_tail/triton/main.py,ok,rel edge_states=1.27e-01 query_projection=3.43e-02 neighbor_projection=3.28e-02 edge_weight=2.17e-02 hidden_weight=2.10e-02 hidden_bias=2.66e-02 output_weight=5.13e-03 output_bias=1.79e-03 norm_weight=4.02e-03 norm_bias=2.93e-03 (band 2e-01 declared),bf16
+mpnn_edge_tail_bwd_dx_recompute_triton,triton,mpnn_edge_tail,miniworld_engine/kernels/mpnn_edge_tail/triton/main.py,ok,rel edge_states=1.27e-01 query_projection=3.43e-02 neighbor_projection=3.28e-02 edge_weight=2.17e-02 hidden_weight=2.10e-02 hidden_bias=2.66e-02 output_weight=5.13e-03 output_bias=1.79e-03 norm_weight=4.02e-03 norm_bias=2.93e-03 (band 2e-01 declared),bf16
 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_node_message_bwd_recompute_triton,triton,mpnn_node_message,miniworld_engine/kernels/mpnn_node_message/triton/main.py,ok,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 1e-01 declared),bf16
+mpnn_node_message_bwd_dx_triton,triton,mpnn_node_message,miniworld_engine/kernels/mpnn_node_message/triton/main.py,ok,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 1e-01 declared),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

{miniworld_engine-2.4.2 → miniworld_engine-2.4.3}/src/miniworld_engine/build/repo_graph.py

@@ -450,6 +450,18 @@ def cuda_rebuilds(result: dict, manifests: list[str]) -> list[tuple[str, str]]:
     return sorted(set(out))
 
 
+#: The card the engine is derived and released for. The derivation check judges THIS card's plan (``autotune/plans/<arch>/``) unless
+#: ``--gpu`` names another: a plan is valid per card (``dev derive --arch X`` re-records only X), so reading a different card's plan
+#: right after deriving this one reports that other plan's stale units as if the derive had not run -- all of them when a file every
+#: unit lists (the safety set: ``settings.py``, ``modules/dispatch.py``, the registry CSVs ...) changed ("1156 of 1156 stale").
+DERIVED_ARCH = "sm90"
+
+
+def derivation_plans() -> dict[str, Path]:
+    """``{arch: units.json}`` of every committed per-card plan."""
+    return {p.parent.name: p for p in sorted((PKG / "autotune" / "plans").glob("*/registry_kernel.units.json"))}
+
+
 def live_evidence(result: dict, new_only: bool = False, gpu: str = "") -> dict:
     """What the committed evidence says NOW, against the working tree (CPU only): the stale tuned caches among the ones reached, the
     native ops whose cache no longer matches their graph, and the derivation units whose recorded files moved.
@@ -466,9 +478,12 @@ def live_evidence(result: dict, new_only: bool = False, gpu: str = "") -> dict:
     reached_ops = {PurePosixPath(n).parent.name for n in result["reached"] if g.nodes[n] == "cache"}
     touched = {p.removeprefix(f"{SRC}/") for p in [*result["changed"], *result["cosmetic"]] if p.startswith(f"{SRC}/")}
     native_stale = sorted({r.op for r in rows if r.stale and r.reason.startswith("native implementation changed")})
-    plan_file = PKG / "autotune" / "plans" / gpu / "registry_kernel.units.json" if gpu else None
-    status = plan.status(plan_file) if plan_file is not None and plan_file.is_file() else plan.status()
+    plans = derivation_plans()
+    arch = gpu or DERIVED_ARCH
+    status = plan.status(plans[arch]) if arch in plans else plan.status()
     units_stale = status["stale"]
+    # the other cards' plans: counted for the report, never judged (a derive on one card does not refresh them)
+    other = {a: len(plan.status(p)["stale"]) for a, p in plans.items() if a != arch}
     if new_only:
         rows_changed = set(result.get("rows") or [])
         seeds = {*result["changed"], *(f"registry-row:{k}" for k in rows_changed)}
@@ -477,7 +492,8 @@ def live_evidence(result: dict, new_only: bool = False, gpu: str = "") -> dict:
         units_stale = {k: v for k, v in units_stale.items() if set(v) & touched}
         reached_ops &= direct                              # a cache whose own files (kernel, helpers, driver, grid, row) hold a changed file
     return {"stale": [r for r in rows if r.stale and r.op in reached_ops], "native_stale": native_stale,
-            "units_stale": units_stale, "units_total": status["units"], "schema": status["schema"], "new_only": new_only}
+            "units_stale": units_stale, "units_total": status["units"], "schema": status["schema"], "new_only": new_only,
+            "units_arch": arch if arch in plans else "", "units_other": other}
 
 
 def _chain(parent: dict[str, str], node: str) -> str:
@@ -501,7 +517,8 @@ def actions(result: dict) -> list[str]:
                     + ", ".join(sorted(native_moved)[:8]) + " -- moved by "
                     + ", ".join(sorted({PurePosixPath(p).name for fs in native_moved.values() for p in fs})[:4]))
     if moved.get(DERIVATION):
-        todo.append(f"dev derive --arch sm86: re-runs only the units whose recorded files changed, reuses the rest; then viz.sweep_page "
+        todo.append(f"dev derive --arch {DERIVED_ARCH} --out src/miniworld_engine/autotune/plans/{DERIVED_ARCH}/registry_kernel.csv: re-runs only the units "
+                    f"whose recorded files changed, reuses the rest; then viz.sweep_page "
                     f"-- moved by {len(moved[DERIVATION])} file(s): " + ", ".join(moved[DERIVATION][:4]) + (" ..." if len(moved[DERIVATION]) > 4 else ""))
     if result.get("rows"):
         todo.append(f"registry rows changed ({len(result['rows'])}): " + ", ".join(result["rows"][:8]) + (" ..." if len(result["rows"]) > 8 else "")
@@ -572,7 +589,11 @@ def report(base: str, verbose: bool = False, manifests: list[str] = (), new_only
     lines.append("owed:")
     lines.extend(f"  - {a}" for a in actions(result) or ["nothing"])
     ev = result["evidence"]
-    lines.append(f"evidence{' (new-only: what this diff causes)' if new_only else ''}{f' [{gpu} only]' if gpu else ''}: derivation {len(ev['units_stale'])} of {ev['units_total']} units stale, "
+    arch_tag = " [" + ev["units_arch"] + "]" if ev.get("units_arch") else ""
+    others = ", ".join(a + " " + str(n) for a, n in sorted(ev.get("units_other", {}).items()))
+    not_judged = f" (not judged: {others} stale)" if others else ""
+    lines.append(f"evidence{' (new-only: what this diff causes)' if new_only else ''}{f' [{gpu} only]' if gpu else ''}: derivation{arch_tag} "
+                 f"{len(ev['units_stale'])} of {ev['units_total']} units stale{not_judged}, "
                  f"native {len(ev['native_stale'])} op(s) stale, tuned {len(ev['stale'])} reached cache(s) stale")
     owed = pending(result)
     lines.append(f"check: {len(owed)} pending" + ("" if not owed else " -- " + "; ".join(owed[:4]) + (" ..." if len(owed) > 4 else "")))

{miniworld_engine-2.4.2 → miniworld_engine-2.4.3}/src/miniworld_engine/kernels/conditioned_transition/cuda/swiglu_ffn.py

@@ -50,6 +50,9 @@ class _SwiGLUFFN(torch.autograd.Function):
 
 _GATE_ONE: dict = {}
 
+#: Widest hidden (rows of ``wa``) at which bf16 inference takes the fused ``ct_inference`` kernel; wider ones take cuBLAS + the SwiGLU row kernel.
+_CT_MAX_HIDDEN_BF16 = 512
+
 
 def _gate_of_one(x: torch.Tensor):
     """(Wsc, bias) of a gate that is exactly one for any cond: zero weights and bias 30 (sigmoid(30) rounds to 1.0f). Cached per (device, dtype,
@@ -77,7 +80,11 @@ def swiglu_ffn(x, wa, wb, ws):
     from miniworld_engine.kernels.conditioned_transition.cuda import ct_inference as ct
 
     d = shape[-1]
-    if _tf32_or_bf16(x) and ct.supported(x2, x2, wa.shape[0], d):
+    # bf16 with a hidden width above 512: the fused expand + SwiGLU GEMM writes a hidden activation that is no longer small next to the GEMM, and
+    # cuBLAS + the SwiGLU row kernel is as fast or faster (CUDA graph, H100, hidden 768 / 1536, 128 - 65536 rows: 0.8 - 1.0x of ct_inference's time;
+    # the fused kernel was 1.04 - 1.37x SLOWER than the PyTorch statements at hidden 1536). At hidden <= 512 the fused kernel stays (1.2 - 1.8x faster).
+    wide_bf16 = x.dtype is torch.bfloat16 and wa.shape[0] > _CT_MAX_HIDDEN_BF16
+    if not wide_bf16 and _tf32_or_bf16(x) and ct.supported(x2, x2, wa.shape[0], d):
         wsc, bsc = _gate_of_one(x2)
         return ct.ct_inference(x2, x2, wa, wb, ws, wsc, bsc).reshape(*shape[:-1], ws.shape[0])
     return _SwiGLUFFN.apply(x2, wa, wb, ws).reshape(*shape[:-1], ws.shape[0])

{miniworld_engine-2.4.2 → miniworld_engine-2.4.3}/src/miniworld_engine/kernels/registry/registry.csv

@@ -112,16 +112,16 @@ transition_squeeze_residual_triton,triton,transition,gemm,both,both,N,both,yes,b
 mpnn_edge_tail_fwd_gemm_gather_saveact_triton,triton,mpnn_edge_tail,gemm,both,both,,mpnn,yes,bf16,miniworld_engine/kernels/mpnn_edge_tail/triton/compute.py,_project_edge,miniworld_engine.kernels.drivers.mpnn_edge_tail:mpnn_edge_tail_fwd_gemm_gather_saveact_triton,miniworld_engine.kernels.checks.mpnn_edge_tail:mpnn_edge_tail_fwd_gemm_gather_saveact_triton,,sm80,,edge,1
 mpnn_edge_tail_fwd_gemm_b2b_saveact_triton,triton,mpnn_edge_tail,gemm,both,both,,mpnn,yes,bf16,miniworld_engine/kernels/mpnn_edge_tail/triton/compute.py,_project_hidden,miniworld_engine.kernels.drivers.mpnn_edge_tail:mpnn_edge_tail_fwd_gemm_b2b_saveact_triton,miniworld_engine.kernels.checks.mpnn_edge_tail:mpnn_edge_tail_fwd_gemm_b2b_saveact_triton,,sm80,,edge,1
 mpnn_edge_tail_fwd_gemm_layernorm_saveact_triton,triton,mpnn_edge_tail,gemm,both,both,,mpnn,yes,bf16,miniworld_engine/kernels/mpnn_edge_tail/triton/compute.py,_project_output,miniworld_engine.kernels.drivers.mpnn_edge_tail:mpnn_edge_tail_fwd_gemm_layernorm_saveact_triton,miniworld_engine.kernels.checks.mpnn_edge_tail:mpnn_edge_tail_fwd_gemm_layernorm_saveact_triton,,sm80,,edge,1
-mpnn_edge_tail_bwd_layernorm_saveact_triton,triton,mpnn_edge_tail,reduce,both,both,,mpnn,yes,bf16,miniworld_engine/kernels/mpnn_edge_tail/triton/compute.py,_norm_backward,miniworld_engine.kernels.drivers.mpnn_edge_tail:mpnn_edge_tail_bwd_layernorm_saveact_triton,miniworld_engine.kernels.checks.mpnn_edge_tail:mpnn_edge_tail_bwd_layernorm_saveact_triton,,sm80,,edge,1
-mpnn_edge_tail_bwd_dx_saveact_triton,triton,mpnn_edge_tail,gemm,both,both,,mpnn,yes,bf16,miniworld_engine/kernels/mpnn_edge_tail/triton/compute.py,_project_backward,miniworld_engine.kernels.drivers.mpnn_edge_tail:mpnn_edge_tail_bwd_dx_saveact_triton,miniworld_engine.kernels.checks.mpnn_edge_tail:mpnn_edge_tail_bwd_dx_saveact_triton,,sm80,,edge,1
-mpnn_edge_tail_bwd_dx_gather_saveact_triton,triton,mpnn_edge_tail,gemm,both,both,,mpnn,yes,bf16,miniworld_engine/kernels/mpnn_edge_tail/triton/compute.py,_edge_backward,miniworld_engine.kernels.drivers.mpnn_edge_tail:mpnn_edge_tail_bwd_dx_gather_saveact_triton,miniworld_engine.kernels.checks.mpnn_edge_tail:mpnn_edge_tail_bwd_dx_gather_saveact_triton,,sm80,,edge,1
+mpnn_edge_tail_bwd_layernorm_saveact_triton,triton,mpnn_edge_tail,reduce,both,both,,mpnn,yes,bf16,miniworld_engine/kernels/mpnn_edge_tail/triton/compute.py,_norm_backward,miniworld_engine.kernels.drivers.mpnn_edge_tail:mpnn_edge_tail_bwd_layernorm_saveact_triton,miniworld_engine.kernels.checks.mpnn_edge_tail:mpnn_edge_tail_bwd_layernorm_saveact_triton,1.0e-1,sm80,,edge,1
+mpnn_edge_tail_bwd_dx_saveact_triton,triton,mpnn_edge_tail,gemm,both,both,,mpnn,yes,bf16,miniworld_engine/kernels/mpnn_edge_tail/triton/compute.py,_project_backward,miniworld_engine.kernels.drivers.mpnn_edge_tail:mpnn_edge_tail_bwd_dx_saveact_triton,miniworld_engine.kernels.checks.mpnn_edge_tail:mpnn_edge_tail_bwd_dx_saveact_triton,1.0e-1,sm80,,edge,1
+mpnn_edge_tail_bwd_dx_gather_saveact_triton,triton,mpnn_edge_tail,gemm,both,both,,mpnn,yes,bf16,miniworld_engine/kernels/mpnn_edge_tail/triton/compute.py,_edge_backward,miniworld_engine.kernels.drivers.mpnn_edge_tail:mpnn_edge_tail_bwd_dx_gather_saveact_triton,miniworld_engine.kernels.checks.mpnn_edge_tail:mpnn_edge_tail_bwd_dx_gather_saveact_triton,1.0e-1,sm80,,edge,1
 mpnn_edge_tail_fwd_gemm_recompute_triton,triton,mpnn_edge_tail,gemm,both,both,,mpnn,yes,bf16,miniworld_engine/kernels/mpnn_edge_tail/triton/main.py,_edge_tail_project_kernel,miniworld_engine.kernels.drivers.mpnn_edge_tail:mpnn_edge_tail_fwd_gemm_recompute_triton,miniworld_engine.kernels.checks.mpnn_edge_tail:mpnn_edge_tail_fwd_gemm_recompute_triton,,sm80,,edge,1
 mpnn_edge_tail_fwd_gemm_layernorm_recompute_triton,triton,mpnn_edge_tail,gemm,both,both,,mpnn,yes,bf16,miniworld_engine/kernels/mpnn_edge_tail/triton/main.py,_edge_tail_norm_kernel,miniworld_engine.kernels.drivers.mpnn_edge_tail:mpnn_edge_tail_fwd_gemm_layernorm_recompute_triton,miniworld_engine.kernels.checks.mpnn_edge_tail:mpnn_edge_tail_fwd_gemm_layernorm_recompute_triton,,sm80,,edge,1
-mpnn_edge_tail_bwd_recompute_triton,triton,mpnn_edge_tail,gemm,both,both,,mpnn,yes,bf16,miniworld_engine/kernels/mpnn_edge_tail/triton/main.py,_edge_tail_replay_kernel,miniworld_engine.kernels.drivers.mpnn_edge_tail:mpnn_edge_tail_bwd_recompute_triton,miniworld_engine.kernels.checks.mpnn_edge_tail:mpnn_edge_tail_bwd_recompute_triton,,sm80,,edge,1
-mpnn_edge_tail_bwd_dx_recompute_triton,triton,mpnn_edge_tail,gemm,both,both,,mpnn,yes,bf16,miniworld_engine/kernels/mpnn_edge_tail/triton/main.py,_edge_tail_dx_kernel,miniworld_engine.kernels.drivers.mpnn_edge_tail:mpnn_edge_tail_bwd_dx_recompute_triton,miniworld_engine.kernels.checks.mpnn_edge_tail:mpnn_edge_tail_bwd_dx_recompute_triton,,sm80,,edge,1
+mpnn_edge_tail_bwd_recompute_triton,triton,mpnn_edge_tail,gemm,both,both,,mpnn,yes,bf16,miniworld_engine/kernels/mpnn_edge_tail/triton/main.py,_edge_tail_replay_kernel,miniworld_engine.kernels.drivers.mpnn_edge_tail:mpnn_edge_tail_bwd_recompute_triton,miniworld_engine.kernels.checks.mpnn_edge_tail:mpnn_edge_tail_bwd_recompute_triton,1.6e-1,sm80,,edge,1
+mpnn_edge_tail_bwd_dx_recompute_triton,triton,mpnn_edge_tail,gemm,both,both,,mpnn,yes,bf16,miniworld_engine/kernels/mpnn_edge_tail/triton/main.py,_edge_tail_dx_kernel,miniworld_engine.kernels.drivers.mpnn_edge_tail:mpnn_edge_tail_bwd_dx_recompute_triton,miniworld_engine.kernels.checks.mpnn_edge_tail:mpnn_edge_tail_bwd_dx_recompute_triton,1.6e-1,sm80,,edge,1
 mpnn_node_message_fwd_gemm_triton,triton,mpnn_node_message,gemm,both,both,,mpnn,yes,bf16,miniworld_engine/kernels/mpnn_node_message/triton/main.py,_node_message_fwd_kernel,miniworld_engine.kernels.drivers.mpnn_node_message:mpnn_node_message_fwd_gemm_triton,miniworld_engine.kernels.checks.mpnn_node_message:mpnn_node_message_fwd_gemm_triton,,sm80,,edge,1
-mpnn_node_message_bwd_recompute_triton,triton,mpnn_node_message,gemm,both,both,,mpnn,yes,bf16,miniworld_engine/kernels/mpnn_node_message/triton/main.py,_node_message_replay_kernel,miniworld_engine.kernels.drivers.mpnn_node_message:mpnn_node_message_bwd_recompute_triton,miniworld_engine.kernels.checks.mpnn_node_message:mpnn_node_message_bwd_recompute_triton,,sm80,,edge,1
-mpnn_node_message_bwd_dx_triton,triton,mpnn_node_message,gemm,both,both,,mpnn,yes,bf16,miniworld_engine/kernels/mpnn_node_message/triton/main.py,_node_message_dx_kernel,miniworld_engine.kernels.drivers.mpnn_node_message:mpnn_node_message_bwd_dx_triton,miniworld_engine.kernels.checks.mpnn_node_message:mpnn_node_message_bwd_dx_triton,,sm80,,edge,1
+mpnn_node_message_bwd_recompute_triton,triton,mpnn_node_message,gemm,both,both,,mpnn,yes,bf16,miniworld_engine/kernels/mpnn_node_message/triton/main.py,_node_message_replay_kernel,miniworld_engine.kernels.drivers.mpnn_node_message:mpnn_node_message_bwd_recompute_triton,miniworld_engine.kernels.checks.mpnn_node_message:mpnn_node_message_bwd_recompute_triton,1.4e-1,sm80,,edge,1
+mpnn_node_message_bwd_dx_triton,triton,mpnn_node_message,gemm,both,both,,mpnn,yes,bf16,miniworld_engine/kernels/mpnn_node_message/triton/main.py,_node_message_dx_kernel,miniworld_engine.kernels.drivers.mpnn_node_message:mpnn_node_message_bwd_dx_triton,miniworld_engine.kernels.checks.mpnn_node_message:mpnn_node_message_bwd_dx_triton,1.4e-1,sm80,,edge,1
 mpnn_relative_position_bwd_reduce_triton,triton,mpnn_relative_position,gemm,both,both,,mpnn,yes,bf16,miniworld_engine/kernels/mpnn_relative_position/triton/main.py,_bucket_reduce_kernel,miniworld_engine.kernels.drivers.mpnn_relative_position:mpnn_relative_position_bwd_reduce_triton,miniworld_engine.kernels.checks.mpnn_relative_position:mpnn_relative_position_bwd_reduce_triton,,sm80,,edge,1
 mpnn_message_fwd_gemm_triton,triton,mpnn_message,gemm,both,both,,mpnn,yes,bf16,miniworld_engine/kernels/mpnn_message/triton/main.py,_projection_fwd_kernel,miniworld_engine.kernels.drivers.mpnn_message:mpnn_message_fwd_gemm_triton,miniworld_engine.kernels.checks.mpnn_message:mpnn_message_fwd_gemm_triton,,sm80,,edge,1
 mpnn_message_fwd_gelu_reduce_triton,triton,mpnn_message,reduce,both,both,,mpnn,yes,bf16,miniworld_engine/kernels/mpnn_message/triton/main.py,_gelu_reduce_fwd_kernel,miniworld_engine.kernels.drivers.mpnn_message:mpnn_message_fwd_gelu_reduce_triton,miniworld_engine.kernels.checks.mpnn_message:mpnn_message_fwd_gelu_reduce_triton,,sm80,,edge,1