Precision Issue Occurs When valid_shape Is Not Passed to view

Symptom

In some scenarios, precision issues occur when valid_shape is not passed to view.

Possible Causes

When the input tensor of the view API does not have a correct validShape, the output validShape cannot be correctly inferred by the framework.

Solution

If you suspect that the validShape inference for view is incorrect, first pass a valid_shape to view and check whether the output meets expectations.

A typical scenario where valid_shape must be passed:

When the input validShape depends on another tensor identifier, dynValidShape must be passed. In the following scenario, the validShape curSeq of q0 comes from another tensor and cannot be obtained through inference:

# Input: input [B, S, H]
# Input: act_seqs [B]
# Output: out [B, S, H]
# Computation process: AddS
# The code is as follows:
for b_idx in pypto.loop(B, name="b_loop", idx_name="b"):
    cur_seq = act_seqs[b_idx]
    a0 = pypto.view(input, [1, S, H], [b_idx, 0, 0], valid_shape=[1, cur_seq, H])
    a1 = a0 + 1.0
    out[b_idx:, :, :] = a1