diff --git a/benchmarks/e2e_decoder.py b/benchmarks/e2e_decoder.py
index daf8241..170be32 100644
--- a/benchmarks/e2e_decoder.py
+++ b/benchmarks/e2e_decoder.py
@@ -15,6 +15,7 @@ Run:
     python benchmarks/e2e_decoder.py
 """
 import os
+import sys
 import time
 import math
 
@@ -22,6 +23,11 @@ os.environ.setdefault("SPARSE_CONV_BACKEND", "flex_gemm")
 os.environ.setdefault("SPARSE_ATTN_BACKEND", "flex_gemm_sparse_attn")
 os.environ.setdefault("FLEX_GEMM_QUIET", "1")
 
+# trellis2 is in-tree (no pip install) — make the bench runnable as
+# `python benchmarks/e2e_decoder.py` regardless of cwd by putting the
+# repo root on sys.path before any trellis2 import.
+sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
+
 import torch
 
 assert torch.backends.mps.is_available(), "This benchmark needs MPS."
diff --git a/trellis2/modules/sparse/config.py b/trellis2/modules/sparse/config.py
index fc3c18d..303f026 100644
--- a/trellis2/modules/sparse/config.py
+++ b/trellis2/modules/sparse/config.py
@@ -10,11 +10,17 @@ def __detect_defaults():
     """Auto-detect best backends for current platform."""
     global CONV, ATTN
     if platform.system() == 'Darwin':
-        ATTN = 'sdpa'
         if __flex_gemm_works_on_mps():
             CONV = 'flex_gemm'
+            # flex_gemm_sparse_attn is now flash-attention-v2 with simdgroup
+            # matmul + simd-shuffle softmax (5–15× over SDPA-padded-CPU-bounce
+            # at every measured shape including max_seqlen=2048). The conv
+            # probe above covers the same package, so if it passed, the
+            # attention path is available too.
+            ATTN = 'flex_gemm_sparse_attn'
         else:
             CONV = 'pytorch'
+            ATTN = 'sdpa'
     elif not __has_cuda():
         CONV = 'pytorch'
         ATTN = 'sdpa'
@@ -97,6 +103,6 @@ def set_debug(debug: bool):
     global DEBUG
     DEBUG = debug
 
-def set_attn_backend(backend: Literal['xformers', 'flash_attn', 'sdpa']):
+def set_attn_backend(backend: Literal['xformers', 'flash_attn', 'flash_attn_3', 'sdpa', 'flex_gemm_sparse_attn']):
     global ATTN
     ATTN = backend