[cherry-pick][0.9.1] vllm-ascend support chunked prefill (vllm-project#1240)

vllm-ascend support chunked prefill for MLA
main 关联pr:vllm-project#1172

---------

<!--  Thanks for sending a pull request!

BEFORE SUBMITTING, PLEASE READ
https://docs.vllm.ai/en/latest/contributing/overview.html

-->
<!--
- Please clarify what changes you are proposing. The purpose of this
section is to outline the changes and how this PR fixes the issue.
If possible, please consider writing useful notes for better and faster
reviews in your PR.

- Please clarify why the changes are needed. For instance, the use case
and bug description.

- Fixes #
-->

<!--
Note that it means *any* user-facing change including all aspects such
as API, interface or other behavior changes.
Documentation-only updates are not considered user-facing changes.
-->

<!--
CI passed with new added/existing test.
If it was tested in a way different from regular unit tests, please
clarify how you tested step by step, ideally copy and paste-able, so
that other reviewers can test and check, and descendants can verify in
the future.
If tests were not added, please describe why they were not added and/or
why it was difficult to add.
-->

Signed-off-by: fems14 <1804143737@qq.com>

Author: fems14

tests/e2e/singlecard/test_chunked.py renamed to tests/singlecard/test_chunked.py

@@ -730,6 +730,83 @@ def _compute_prefill_context(
                 softmax_lse=prefix_lse)
         return prefix_output, prefix_lse
 
+    def _compute_prefill_context(
+        self,
+        query: torch.Tensor,
+        kv_c_and_k_pe_cache: torch.Tensor,
+        rope_dim: int,
+        attn_metadata: AscendMLAMetadata,
+        prefix_output: torch.Tensor,
+        prefix_lse: torch.Tensor,
+    ):
+        prefill_metadata = attn_metadata.prefill
+        if prefill_metadata is None or prefill_metadata.chunked_context is None:
+            return prefix_output, prefix_lse
+
+        iters = len(prefill_metadata.chunked_context.seq_tot)
+        q_pe = query[..., self.qk_nope_head_dim:]
+        q_nope = query[..., :self.qk_nope_head_dim]
+
+        seq_len1 = torch.tensor(prefill_metadata.query_lens, dtype=torch.int32)
+        latent_kv_dim = kv_c_and_k_pe_cache.size(3) - rope_dim
+        cache_kv_c = kv_c_and_k_pe_cache[:, :, :, :latent_kv_dim]
+        cache_k_pe = kv_c_and_k_pe_cache[:, :, :, latent_kv_dim:]
+        for i in range(iters):
+            toks = prefill_metadata.chunked_context.seq_tot[i]
+
+            seq_len2 = prefill_metadata.chunked_context.chunk_seq_lens[i]
+            seq_len = torch.stack([seq_len1, seq_len2])
+            kv_c_normed = torch.empty(toks,
+                                      kv_c_and_k_pe_cache.size(2),
+                                      latent_kv_dim,
+                                      dtype=query.dtype,
+                                      device=query.device)
+            k_pe = torch.empty(toks,
+                               kv_c_and_k_pe_cache.size(2),
+                               rope_dim,
+                               dtype=query.dtype,
+                               device=query.device)
+
+            torch_npu.atb.npu_paged_cache_load(
+                cache_kv_c,
+                cache_k_pe,
+                prefill_metadata.block_table,
+                seq_len2.to(query.device),
+                seq_starts=prefill_metadata.chunked_context.starts[i],
+                key=kv_c_normed,
+                value=k_pe,
+            )
+
+            kv_c_normed = kv_c_normed.squeeze()
+            kv_nope = self.kv_b_proj(kv_c_normed)[0].view( \
+                -1, self.num_heads, self.qk_nope_head_dim + self.v_head_dim)
+            k_nope, v = kv_nope\
+                .split([self.qk_nope_head_dim, self.v_head_dim], dim=-1)
+            k_pe = k_pe.expand((*k_nope.shape[:-1], -1))
+            mask = torch.triu(
+                torch.ones(512, 512, device=query.device, dtype=query.dtype),
+                1)
+            torch_npu.atb.npu_ring_mla(
+                q_nope=q_nope,
+                q_rope=q_pe,
+                k_nope=k_nope,
+                k_rope=k_pe,
+                value=v,
+                mask=mask,
+                seqlen=seq_len,
+                head_num=self.num_heads,
+                kv_head_num=self.num_heads,
+                pre_out=prefix_output,
+                prev_lse=prefix_lse,
+                qk_scale=self.scale,
+                kernel_type="kernel_type_high_precision",
+                mask_type="no_mask",
+                input_layout="type_bsnd",
+                calc_type="calc_type_default",
+                output=prefix_output,
+                softmax_lse=prefix_lse)
+        return prefix_output, prefix_lse
+
     def _forward_prefill(
         self,
         query: torch.Tensor,