fix: errors caused by trivector storage removal

turboquant/benchmark_google_parity.py

@@ -66,7 +66,7 @@ def compress(ks, li):
         rq, pk = compressors[li]
         flat = ks.reshape(-1, D).float()
         kq = triton_rotor_full_fused(flat, pk, None,
-            getattr(rq, 'centroids_vector'), None, getattr(rq, 'centroids_trivector'))
+            getattr(rq, 'centroids_vector'), None, None)
         return kq.to(ks.dtype).reshape(ks.shape)
 
     _orig = DynamicCache.update

turboquant/benchmark_perplexity.py

@@ -94,7 +94,7 @@ def compress(ks, li):
             flat, pk, None,
             getattr(rq, 'centroids_vector'),
             None,
-            getattr(rq, 'centroids_trivector'),
+            None,
         )
         return kq.to(ks.dtype).reshape(ks.shape)
 

turboquant/benchmark_triton.py

@@ -114,7 +114,7 @@ def verify_correctness():
     c_scalar = None
     c_vector = getattr(rq, 'centroids_vector')
     c_bivector = None
-    c_trivector = getattr(rq, 'centroids_trivector')
+    c_trivector = None
 
     x_hat_triton = triton_rotor_full_fused(
         x, packed_rotors, c_scalar, c_vector, c_bivector, c_trivector)
@@ -197,7 +197,7 @@ def benchmark_full_fused():
     c_s = None
     c_v = getattr(rq, 'centroids_vector')
     c_b = None
-    c_t = getattr(rq, 'centroids_trivector')
+    c_t = None
 
     print(f"  GPU: {torch.cuda.get_device_name()}")
     print(f"  d={d}, bits={bits}\n")
@@ -328,7 +328,7 @@ def benchmark_varying_dimensions():
         c_s = None
         c_v = getattr(rq, 'centroids_vector')
         c_b = None
-        c_t = getattr(rq, 'centroids_trivector')
+        c_t = None
         n_groups = (d + 2) // 3
 
         x = torch.randn(n, d, device=device)
@@ -381,7 +381,7 @@ def benchmark_vs_turboquant():
         c_s = None
         c_v = getattr(rq, 'centroids_vector')
         c_b = None
-        c_t = getattr(rq, 'centroids_trivector')
+        c_t = None
 
         for n in [1024, 4096, 16384]:
             x = torch.randn(n, d, device=device)
@@ -437,7 +437,7 @@ def benchmark_bitwidth_sweep():
         c_s = None
         c_v = getattr(rq, 'centroids_vector')
         c_b = None
-        c_t = getattr(rq, 'centroids_trivector')
+        c_t = None
 
         # Quality
         x_hat_pt, _ = rq(x)

turboquant/poc_high_context.py

@@ -37,7 +37,6 @@ def __init__(self, head_dim: int, bits: int, seed: int, device: str):
         self.rq = RotorQuantMSE(head_dim, bits, seed=seed, device=device)
         self.packed_rotors = pack_rotors_for_triton(self.rq.rotors).to(device)
         self.c_v = getattr(self.rq, 'centroids_vector').to(device)
-        self.c_t = getattr(self.rq, 'centroids_trivector').to(device)
         self.head_dim = head_dim
         self.device = device
 
@@ -59,7 +58,7 @@ def compress_dequantize(self, keys: torch.Tensor) -> torch.Tensor:
         # Triton fused: embed→rotor→quantize→unrotor→extract
         flat_recon = triton_rotor_full_fused(
             flat, self.packed_rotors,
-            None, self.c_v, None, self.c_t,
+            None, self.c_v, None, None,
         )
 
         return flat_recon.to(orig_dtype).reshape(B, H, S, D)

turboquant/triton_kernels.py

@@ -289,7 +289,7 @@ def triton_rotor_full_fused(
     c_s = c_scalar.float().contiguous() if c_scalar is not None else c_vector.float().contiguous()
     c_v = c_vector.float().contiguous()
     c_b = c_bivector.float().contiguous() if c_bivector is not None else c_vector.float().contiguous()
-    c_t = c_trivector.float().contiguous()
+    c_t = c_trivector.float().contiguous() if c_trivector is not None else c_vector.float().contiguous()
 
     output = torch.empty_like(input_f32)