Verify each finding against the current code and only fix it if needed.

Inline comments:
In `@tests/unittest/_torch/thop/parallel/test_w4a8_mxfp4_mxfp8_gemm.py`:
- Line 49: The division by mat_b.abs().max().float() in the global_scale_b
calculation can produce inf for zero tensors; guard the denominator by computing
a safe max (e.g., max_val = mat_b.abs().max().float() and then replace zero with
a small epsilon or use torch.clamp/max to enforce a minimum like 1e-6) and use
that safe_max in the division so global_scale_b = (448 * 6) / safe_max; update
references to mat_b.abs().max().float() in this expression to the safe_max
variable.
- Around line 58-60: The test currently only checks cosine similarity
(scale-invariant) and can miss cases where alpha is ignored; update the
assertion after computing c and c_ref (variables: c, c_ref, alpha, mat_a, mat_b)
to include a magnitude-sensitive check — e.g., assert that the tensor norms or
element-wise values match the expected scaled result using torch.allclose or a
max-abs-difference threshold (for example compare c.norm() to c_ref.norm() with
a small rtol/atol or assert torch.allclose(c, c_ref, rtol=1e-3, atol=1e-6)) so
the test fails if alpha is not applied correctly.