fix(glm53): split-K head gate 가 block_k 로 K 를 못 나누면 조용한 오답 대신 실패

build/glm53/glm53_indexer_gate.py

@@ -86,6 +86,11 @@ def head_gate_splitk(x: torch.Tensor, w: torch.Tensor, block_k: int = _BLOCK_K)
     fp32, contiguous). Returns a fresh contiguous fp32 [M, N] tensor."""
     M, K = x.shape
     N = w.shape[1]
+    if K % block_k:
+        # The partial kernel reads w rows unmasked over [0, split*block_k): a K
+        # the block does not tile would drop the tail and answer quietly wrong.
+        raise ValueError(
+            f"head_gate_splitk: K={K} is not a multiple of block_k={block_k}")
     bn = _bn_for(N)
     split = K // block_k
     part = torch.empty(split * MAX_M, bn, device=x.device, dtype=torch.float32)

overlay/modules/glm53_indexer_gate_splitk/glm53_indexer_gate.py

@@ -86,6 +86,11 @@ def head_gate_splitk(x: torch.Tensor, w: torch.Tensor, block_k: int = _BLOCK_K)
     fp32, contiguous). Returns a fresh contiguous fp32 [M, N] tensor."""
     M, K = x.shape
     N = w.shape[1]
+    if K % block_k:
+        # The partial kernel reads w rows unmasked over [0, split*block_k): a K
+        # the block does not tile would drop the tail and answer quietly wrong.
+        raise ValueError(
+            f"head_gate_splitk: K={K} is not a multiple of block_k={block_k}")
     bn = _bn_for(N)
     split = K // block_k
     part = torch.empty(split * MAX_M, bn, device=x.device, dtype=torch.float32)

tests/test_logic.py

@@ -7184,6 +7184,8 @@ def test_glm53_indexer_gate_splitk_contracts() -> None:
           "applicability guards x's layout and the K match (the kernel assumes both)")
     check("w.dtype == torch.float32" in kern and "w.is_contiguous()" in kern,
           "applicability: fp32 contiguous weight only")
+    check("K % block_k" in kern and "raise ValueError" in kern,
+          "a non-tiling block_k fails loud (the partial kernel reads w rows unmasked)")
     check("return torch.mm(x.float(), w)" in kern, "the helper falls back to the stock product")
     check("[indexer-gate]" in kern and "logger.warning" in kern,
           "routing is announced once per shape so an armed boot that never runs split-K is visible")