在 speculative mode 下，测试 Attention 在传入 mask 下的功能
(self)
| 353 | ) |
| 354 | |
| 355 | def test_mask(self): |
| 356 | """ |
| 357 | 在 speculative mode 下，测试 Attention 在传入 mask 下的功能 |
| 358 | """ |
| 359 | prefill_len = 8192 |
| 360 | dec_len_q = 5 |
| 361 | total_len = prefill_len + dec_len_q |
| 362 | mask = paddle.tril(paddle.ones((self.bsz, dec_len_q, total_len), dtype="float32"), diagonal=prefill_len) |
| 363 | mask_ref = paddle.where(mask == 1, paddle.zeros_like(mask), paddle.full_like(mask, fill_value=float("-inf"))) |
| 364 | |
| 365 | mask_append_attn = mask[:, :, prefill_len:] |
| 366 | mask_append_attn = paddle.where( |
| 367 | mask_append_attn == 1, |
| 368 | paddle.full_like(mask_append_attn, fill_value=False, dtype=bool), |
| 369 | paddle.full_like(mask_append_attn, fill_value=True, dtype=bool), |
| 370 | ) |
| 371 | |
| 372 | self.run_append_c16_attention(prefill_len, 0, True) |
| 373 | dec_out = self.run_append_c16_attention(dec_len_q, prefill_len, False, mask_append_attn) |
| 374 | |
| 375 | ref_out = self.ref_attention(self.CURRENT_Q[0], self.TOTAL_K, self.TOTAL_V, mask_ref) |
| 376 | |
| 377 | np.testing.assert_allclose( |
| 378 | ref_out.astype("float32").numpy(), dec_out.astype("float32").numpy(), rtol=1e-03, atol=5e-03 |
| 379 | ) |
| 380 | |
| 381 | def test_tree_mask(self): |
| 382 | """ |
nothing calls this directly
no test coverage detected