miniworld-engine 2.3.3__tar.gz → 2.3.5__tar.gz

{miniworld_engine-2.3.3 → miniworld_engine-2.3.5}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: miniworld-engine
-Version: 2.3.3
+Version: 2.3.5
 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 2.3.3** — CUDA + Triton only (CuTe/quack removed), torch 2.13 (CUDA 12.9 or 13.0), prebuilt H100 kernels,
+**Version 2.3.5** — 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).
 

{miniworld_engine-2.3.3 → miniworld_engine-2.3.5}/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 2.3.3** — CUDA + Triton only (CuTe/quack removed), torch 2.13 (CUDA 12.9 or 13.0), prebuilt H100 kernels,
+**Version 2.3.5** — 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).
 

{miniworld_engine-2.3.3 → miniworld_engine-2.3.5}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "miniworld-engine"
-version = "2.3.3"
+version = "2.3.5"
 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-2.3.3 → miniworld_engine-2.3.5}/src/miniworld_engine/autotune/manifests/NVIDIA H100 80GB HBM3 (sm90).csv

@@ -1,9 +1,9 @@
 kernel,backend,family,file,status,detail,dtype
-#provenance,2.3.2,230ef3679ec8,clean,2026-10-20,v2.3.1-5-g230ef367,
-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,2.3.4,ca7d86bb8dd4,clean,2026-10-21,v2.3.3-2-gca7d86bb,
+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=4.41e-07 dbeta=2.50e-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=3.46e-07 dbeta=1.67e-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
@@ -45,7 +45,7 @@ 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

{miniworld_engine-2.3.3 → miniworld_engine-2.3.5}/src/miniworld_engine/build/bundle_coverage.py

@@ -13,7 +13,7 @@ the number of batched samples. This driver runs, on the current GPU:
   pair shared by the samples or one per sample, masked, hoisted, d_cond 384 / 768), the SWA atom transformer in its
   three styles and the block-local atom DiT), inference and training,
 
-every one in bf16, fp32, bf16-mixed (fp32 parameters under bf16 autocast, MiniWorld's precision) and, for the blocks,
+every one in bf16, fp32, bf16-mixed (fp32 parameters under bf16 autocast, MiniWorld's precision) and
 fp32 parameters fed bf16 activations (``PRECISIONS``).
 
 The ``kernels/prebuilt`` recorder (``MINIWORLD_ENGINE_BUNDLE_RECORD``) captures what was built or loaded; this module
@@ -75,7 +75,9 @@ def cases() -> list[dict]:
     for name in CASE_NAMES:
         shapes = _shapes(name)
         dtypes = [str(d).removeprefix("torch.") for d in shapes["dtypes"]] or ["bfloat16"]
-        precs = [*dtypes, "mixed"]
+        # "master": fp32 parameters fed bf16 activations (pdall's training, no autocast) selects other kernel builds
+        # (single-direction TriMul backward with -DMASTER_FP32=1); 2.3.4 shipped without them.
+        precs = [*dtypes, "mixed", *(["master"] if "bfloat16" in dtypes else [])]
         computes = [str(d).removeprefix("torch.") for d in shapes["compute_dtypes"]] or [None]
         modes = [False, True] if shapes["train"] else [False]
         for i, (stream, registry) in enumerate(zip(shapes["streams"], shapes["lengths_by_dim"], strict=True)):

{miniworld_engine-2.3.3 → miniworld_engine-2.3.5}/src/miniworld_engine.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: miniworld-engine
-Version: 2.3.3
+Version: 2.3.5
 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 2.3.3** — CUDA + Triton only (CuTe/quack removed), torch 2.13 (CUDA 12.9 or 13.0), prebuilt H100 kernels,
+**Version 2.3.5** — 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).