vllm-cpu-amxbf16 0.11.2.post2__cp310-cp310-manylinux_2_17_x86_64.whl

vllm/v1/attention/backends/mla/indexer.py

@@ -0,0 +1,362 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+from dataclasses import dataclass
+from typing import ClassVar
+
+import torch
+
+from vllm.attention.backends.abstract import (
+    AttentionBackend,
+    MultipleOf,
+)
+from vllm.config import VllmConfig
+from vllm.logger import init_logger
+from vllm.utils.deep_gemm import get_paged_mqa_logits_metadata
+from vllm.v1.attention.backends.utils import (
+    AttentionCGSupport,
+    AttentionMetadataBuilder,
+    CommonAttentionMetadata,
+    split_decodes_and_prefills,
+)
+
+logger = init_logger(__name__)
+
+
+class DeepseekV32IndexerBackend(AttentionBackend):
+    supported_kernel_block_sizes: ClassVar[list[int | MultipleOf]] = [64]
+
+    @classmethod
+    def get_supported_head_sizes(cls) -> list[int]:
+        return [32, 64, 128]
+
+    @staticmethod
+    def get_builder_cls() -> type["DeepseekV32IndexerMetadataBuilder"]:
+        return DeepseekV32IndexerMetadataBuilder
+
+    @staticmethod
+    def get_kv_cache_shape(
+        num_blocks: int,
+        block_size: int,
+        num_kv_heads: int,
+        head_size: int,
+        cache_dtype_str: str = "auto",
+    ) -> tuple[int, ...]:
+        assert num_kv_heads == 1
+        return (num_blocks, block_size, head_size)
+
+    @staticmethod
+    def get_kv_cache_stride_order() -> tuple[int, ...]:
+        return (0, 1, 2)
+
+
+@dataclass
+class DeepseekV32IndexerPrefillChunkMetadata:
+    block_table: torch.Tensor
+    cu_seqlen_ks: torch.Tensor
+    cu_seqlen_ke: torch.Tensor
+    cu_seq_lens: torch.Tensor
+    total_seq_lens: int
+    token_start: int
+    token_end: int
+    num_reqs: int
+
+
+@dataclass
+class DeepseekV32IndexerPrefillMetadata:
+    chunks: list[DeepseekV32IndexerPrefillChunkMetadata]
+
+
+@dataclass
+class DeepSeekV32IndexerDecodeMetadata:
+    block_table: torch.Tensor
+    seq_lens: torch.Tensor
+    decode_lens: torch.Tensor
+    requires_padding: bool
+    schedule_metadata: torch.Tensor
+
+
+@dataclass
+class DeepseekV32IndexerMetadata:
+    # FIXME (zyongye)
+    # hacky way to access the data now, need to be in chunked meta
+    seq_lens: torch.Tensor
+
+    num_reqs: int
+    max_query_len: int
+    max_seq_len: int
+
+    num_actual_tokens: int  # Number of tokens excluding padding.
+    query_start_loc: torch.Tensor
+    slot_mapping: torch.Tensor
+    # The dimension of the attention heads
+    head_dim: int
+
+    # New for MLA (compared to FlashAttention)
+    # For handling prefill decode split
+    num_decodes: int
+    num_decode_tokens: int
+    num_prefills: int
+    num_prefill_tokens: int
+
+    decode: DeepSeekV32IndexerDecodeMetadata | None = None
+    prefill: DeepseekV32IndexerPrefillMetadata | None = None
+
+
+# TODO (zyongye) optimize this, this is now vibe coded
+def kv_spans_from_batches(
+    start_seq_loc: torch.Tensor, seq_len_per_batch: torch.Tensor, device: torch.device
+) -> tuple[torch.Tensor, torch.Tensor]:
+    """
+    Args:
+      start_seq_loc: 1D long tensor [B+1], cumulative counts of
+                     selected tokens per batch.
+            Example: [0, 2, 4, 7] ->
+                     batch sizes (selected) [2, 2, 3], N=7 tokens total.
+      seq_len_per_batch: 1D long tensor [B],
+                         full sequence length (KV length) of each batch.
+                         Example: [5, 9, 4].
+
+    Returns:
+      start_tensor: 1D long tensor [N], start offset in the
+                    concatenated KV cache for each token's batch.
+      end_location: 1D long tensor [N],
+                    **exclusive** end = start + token's local position.
+                    (So the attended KV slice is kv[start:end].)
+
+    Assumes each batch contributes its full `seq_len_per_batch[i]`
+    keys to the KV cache, andthe selected tokens within a batch
+    are the **last** `counts[i]` positions of that sequence.
+    """
+    q = start_seq_loc.to(dtype=torch.long)
+    L = seq_len_per_batch.to(dtype=torch.long)
+    assert q.dim() == 1 and L.dim() == 1
+    assert q.numel() == L.numel() + 1, "start_seq_loc must have length B+1"
+
+    # Selected tokens per batch and totals
+    counts = q[1:] - q[:-1]  # [B]
+    N = int(q[-1].item())  # total selected tokens
+    B = L.numel()
+
+    if N == 0:
+        return (
+            torch.empty(0, dtype=torch.long, device=device),
+            torch.empty(0, dtype=torch.long, device=device),
+        )
+
+    # KV start offsets per batch in the concatenated KV cache
+    kv_starts_per_batch = torch.cumsum(L, dim=0) - L  # [B]
+
+    # For each selected token, which batch does it belong to?
+    batch_id = torch.repeat_interleave(torch.arange(B), counts)  # [N]
+
+    # Map batch KV start to each token
+    start_tensor = kv_starts_per_batch[batch_id]  # [N]
+
+    # End-align local positions inside each batch:
+    # local_pos = L[b] - counts[b] + (1..counts[b])  for each batch b
+    L_expand = torch.repeat_interleave(L, counts)  # [N]
+    m_expand = torch.repeat_interleave(counts, counts)  # [N]
+    # position within the selected block: 1..counts[b]
+    pos_within = (
+        torch.arange(N, dtype=torch.long) - torch.repeat_interleave(q[:-1], counts) + 1
+    )
+
+    local_pos = L_expand - m_expand + pos_within  # [N], 1-based
+    end_location = start_tensor + local_pos  # exclusive end
+
+    return start_tensor.int().to(device), end_location.int().to(device)
+
+
+def get_max_prefill_buffer_size(vllm_config: VllmConfig):
+    max_model_len = vllm_config.model_config.max_model_len
+    # NOTE(Chen): 2 is a magic number for controlling the prefill buffer size.
+    # May be tuned later.
+    return max_model_len * 2
+
+
+def split_prefill_chunks(
+    seq_lens_cpu: torch.Tensor, max_prefill_buffer_size: int, reqs_start: int
+) -> list[tuple[int, int]]:
+    """
+    Split the prefill chunks into a list of tuples of (reqs_start, reqs_end)
+    such that the total sequence length of each chunk is less than the
+    maximum prefill buffer size.
+
+    Args:
+        seq_lens_cpu: The sequence lengths of the prefill requests.
+        max_prefill_buffer_size: The maximum prefill buffer size.
+        reqs_start: The start index of the prefill requests.
+
+    Returns:
+        A list of tuples of (reqs_start, reqs_end).
+    """
+    chunk_seq_ids = []
+    total_seq_lens = 0
+    for i in range(reqs_start, len(seq_lens_cpu)):
+        cur_seq_len = seq_lens_cpu[i].item()
+        assert cur_seq_len <= max_prefill_buffer_size
+        total_seq_lens += cur_seq_len
+        if total_seq_lens > max_prefill_buffer_size:
+            chunk_seq_ids.append((reqs_start, i))
+            reqs_start = i
+            total_seq_lens = cur_seq_len
+    if total_seq_lens > 0:
+        chunk_seq_ids.append((reqs_start, len(seq_lens_cpu)))
+    return chunk_seq_ids
+
+
+class DeepseekV32IndexerMetadataBuilder(AttentionMetadataBuilder):
+    _cudagraph_support: ClassVar[AttentionCGSupport] = (
+        AttentionCGSupport.UNIFORM_SINGLE_TOKEN_DECODE
+    )
+
+    reorder_batch_threshold: int = 1
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        scheduler_config = self.vllm_config.scheduler_config
+        # NOTE(Chen):an estimated max size of flattened_kv. Need to double check.
+        self.max_prefill_buffer_size = get_max_prefill_buffer_size(self.vllm_config)
+        self.num_speculative_tokens = (
+            self.vllm_config.speculative_config.num_speculative_tokens
+            if self.vllm_config.speculative_config
+            else 0
+        )
+        # Now deepgemm fp8_paged_mqa_logits does not support next_n > 2
+        self.reorder_batch_threshold += min(self.num_speculative_tokens, 1)
+
+        props = torch.cuda.get_device_properties(self.device)
+        sm_count = props.multi_processor_count
+        self.num_sms = sm_count
+
+        self.decode_lens_buffer = torch.empty(
+            (scheduler_config.max_num_seqs,), dtype=torch.int32, device=self.device
+        )
+
+        # See: DeepGMM/csrc/apis/attention.hpp
+        self.scheduler_metadata_buffer = torch.empty(
+            (self.num_sms + 1, 2), dtype=torch.int32, device=self.device
+        )
+
+    def build_one_prefill_chunk(
+        self, reqs_start, reqs_end, query_start_loc_cpu, seq_lens_cpu, block_table
+    ):
+        prefill_query_start_loc = (
+            query_start_loc_cpu[reqs_start : reqs_end + 1]
+            - query_start_loc_cpu[reqs_start]
+        )
+        cu_seqlen_ks, cu_seqlen_ke = kv_spans_from_batches(
+            prefill_query_start_loc, seq_lens_cpu[reqs_start:reqs_end], self.device
+        )
+        token_start = query_start_loc_cpu[reqs_start].item()
+        token_end = query_start_loc_cpu[reqs_end].item()
+        total_seq_lens = seq_lens_cpu[reqs_start:reqs_end].sum()
+        assert total_seq_lens <= self.max_prefill_buffer_size
+        cu_seq_lens = (
+            torch.cat(
+                [
+                    torch.zeros(1, dtype=torch.int32),
+                    seq_lens_cpu[reqs_start:reqs_end].cumsum(dim=0),
+                ]
+            )
+            .to(torch.int32)
+            .to(self.device)
+        )
+        return DeepseekV32IndexerPrefillChunkMetadata(
+            cu_seqlen_ks=cu_seqlen_ks,
+            cu_seqlen_ke=cu_seqlen_ke,
+            cu_seq_lens=cu_seq_lens,
+            total_seq_lens=total_seq_lens,
+            block_table=block_table[reqs_start:reqs_end],
+            token_start=token_start,
+            token_end=token_end,
+            num_reqs=reqs_end - reqs_start,
+        )
+
+    def build(
+        self,
+        common_prefix_len: int,
+        common_attn_metadata: CommonAttentionMetadata,
+        fast_build: bool = False,
+    ) -> DeepseekV32IndexerMetadata:
+        num_reqs = common_attn_metadata.num_reqs
+        num_tokens = common_attn_metadata.num_actual_tokens
+
+        query_start_loc_cpu = common_attn_metadata.query_start_loc_cpu
+        num_decodes, num_prefills, num_decode_tokens, num_prefill_tokens = (
+            split_decodes_and_prefills(
+                common_attn_metadata, decode_threshold=self.reorder_batch_threshold
+            )
+        )
+
+        assert num_decodes + num_prefills == num_reqs
+        assert num_decode_tokens + num_prefill_tokens == num_tokens
+
+        prefill_metadata = None
+        if num_prefills > 0:
+            chunk_seq_ids = split_prefill_chunks(
+                common_attn_metadata.seq_lens_cpu,
+                self.max_prefill_buffer_size,
+                num_decodes,
+            )
+            chunks = [
+                self.build_one_prefill_chunk(
+                    reqs_start,
+                    reqs_end,
+                    query_start_loc_cpu,
+                    common_attn_metadata.seq_lens_cpu,
+                    common_attn_metadata.block_table_tensor,
+                )
+                for reqs_start, reqs_end in chunk_seq_ids
+            ]
+            prefill_metadata = DeepseekV32IndexerPrefillMetadata(
+                chunks=chunks,
+            )
+
+        decode_metadata = None
+        if num_decodes > 0:
+            torch.diff(
+                common_attn_metadata.query_start_loc[: num_decodes + 1],
+                out=self.decode_lens_buffer[:num_decodes],
+            )
+            decode_lens = self.decode_lens_buffer[:num_decodes]
+            decode_lens_cpu = torch.diff(
+                common_attn_metadata.query_start_loc_cpu[: num_decodes + 1]
+            )
+
+            # Use CPU to avoid GPU sync; breaking async scheduling
+            requires_padding = (decode_lens_cpu.max() > decode_lens_cpu.min()).item()
+
+            seq_lens = common_attn_metadata.seq_lens[:num_decodes]
+
+            self.scheduler_metadata_buffer[:] = get_paged_mqa_logits_metadata(
+                seq_lens, self.kv_cache_spec.block_size, self.num_sms
+            )
+            decode_metadata = DeepSeekV32IndexerDecodeMetadata(
+                block_table=common_attn_metadata.block_table_tensor[:num_decodes, ...],
+                seq_lens=common_attn_metadata.seq_lens[:num_decodes],
+                decode_lens=decode_lens,
+                requires_padding=requires_padding,
+                schedule_metadata=self.scheduler_metadata_buffer,
+            )
+
+        attn_metadata = DeepseekV32IndexerMetadata(
+            seq_lens=common_attn_metadata.seq_lens,
+            num_reqs=common_attn_metadata.num_reqs,
+            max_query_len=common_attn_metadata.max_query_len,
+            max_seq_len=common_attn_metadata.max_seq_len,
+            num_actual_tokens=common_attn_metadata.num_actual_tokens,
+            query_start_loc=common_attn_metadata.query_start_loc,
+            slot_mapping=common_attn_metadata.slot_mapping,
+            head_dim=128,
+            num_decodes=num_decodes,
+            num_decode_tokens=num_decode_tokens,
+            num_prefills=num_prefills,
+            num_prefill_tokens=num_prefill_tokens,
+            prefill=prefill_metadata,
+            decode=decode_metadata,
+        )
+
+        # if get_tensor_model_parallel_rank() == 0:
+        #     logger.info(f"attn_metadata: {attn_metadata}")
+        return attn_metadata

vllm/v1/attention/backends/mla/rocm_aiter_mla.py

@@ -0,0 +1,294 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+from dataclasses import dataclass
+from typing import ClassVar
+
+import torch
+
+from vllm._aiter_ops import rocm_aiter_ops
+from vllm.attention.backends.abstract import AttentionLayer
+from vllm.config import VllmConfig
+from vllm.utils.math_utils import cdiv
+from vllm.v1.attention.backends.mla.common import (
+    MLACommonBackend,
+    MLACommonDecodeMetadata,
+    MLACommonImpl,
+    MLACommonMetadata,
+    MLACommonMetadataBuilder,
+)
+from vllm.v1.attention.backends.utils import AttentionCGSupport
+from vllm.v1.kv_cache_interface import AttentionSpec
+
+
+class AiterMLABackend(MLACommonBackend):
+    @staticmethod
+    def get_name() -> str:
+        return "ROCM_AITER_MLA"
+
+    @staticmethod
+    def get_impl_cls() -> type["AiterMLAImpl"]:
+        return AiterMLAImpl
+
+    @staticmethod
+    def get_builder_cls() -> type["AiterMLAMetadataBuilder"]:
+        return AiterMLAMetadataBuilder
+
+
+@dataclass
+class AiterMLADecodeMetadata(MLACommonDecodeMetadata):
+    # The indptr of the paged kv cache, shape: [batch_size + 1]
+    paged_kv_indptr: torch.Tensor | None = None
+    # The page indices of the paged kv cache
+    paged_kv_indices: torch.Tensor | None = None
+    # The number of entries in the last page of each request in
+    # the paged kv cache, shape: [batch_size]
+    paged_kv_last_page_len: torch.Tensor | None = None
+    # The query indptr, shape : [num_decode + 1]
+    qo_indptr: torch.Tensor | None = None
+
+
+class AiterMLAMetadata(MLACommonMetadata[AiterMLADecodeMetadata]):
+    pass
+
+
+class AiterMLAMetadataBuilder(MLACommonMetadataBuilder[AiterMLAMetadata]):
+    # TODO(luka, lucas): audit this as part of:
+    #  https://github.com/vllm-project/vllm/issues/22945
+    _cudagraph_support: ClassVar[AttentionCGSupport] = (
+        AttentionCGSupport.UNIFORM_SINGLE_TOKEN_DECODE
+    )
+
+    def __init__(
+        self,
+        kv_cache_spec: AttentionSpec,
+        layer_names: list[str],
+        vllm_config: VllmConfig,
+        device: torch.device,
+    ):
+        super().__init__(
+            kv_cache_spec, layer_names, vllm_config, device, AiterMLAMetadata
+        )
+
+        self.compilation_config = vllm_config.compilation_config
+        max_num_pages_per_req = cdiv(
+            vllm_config.model_config.max_model_len, self.kv_cache_spec.block_size
+        )
+        max_num_reqs = vllm_config.scheduler_config.max_num_seqs
+        max_num_pages = max_num_reqs * max_num_pages_per_req
+
+        # Preparing persistent buffers
+        # TODO: we can disambiguate between decode and mixed-prefill decode here
+        # so we can only use the persistent buffer if a cudagraph is actually
+        # being used.
+        if self.compilation_config.cudagraph_mode.has_full_cudagraphs():
+            self.block_table_remapping = torch.zeros(
+                [max_num_reqs, max_num_pages_per_req * self.kv_cache_spec.block_size],
+                dtype=torch.int32,
+                device=device,
+            )
+            self.paged_kv_indptr = torch.zeros(
+                max_num_reqs + 1, dtype=torch.int32, device=device
+            )
+            self.paged_kv_indices = torch.zeros(
+                max_num_pages, dtype=torch.int32, device=device
+            )
+            self.paged_kv_last_page_len = torch.zeros(
+                max_num_reqs, dtype=torch.int32, device=device
+            )
+
+            self.qo_indptr = torch.arange(
+                0, max_num_reqs + 1, dtype=torch.int32, device=device
+            )
+
+    def _build_decode(
+        self,
+        block_table_tensor: torch.Tensor,
+        seq_lens_cpu: torch.Tensor,
+        seq_lens_device: torch.Tensor,
+        query_start_loc_cpu: torch.Tensor,
+        query_start_loc_device: torch.Tensor,
+        num_decode_tokens: int,
+        dcp_tot_seq_lens_device: torch.Tensor | None,
+    ) -> AiterMLADecodeMetadata:
+        page_size = self.kv_cache_spec.block_size
+        device = self.device
+        num_reqs = seq_lens_device.size(0)
+        bs, _ = block_table_tensor.shape
+        block_table_tensor = (
+            block_table_tensor.unsqueeze(-1).expand(-1, -1, page_size) * page_size
+        )
+        block_table_tensor = (
+            block_table_tensor
+            + torch.arange(
+                0,
+                page_size,
+                device=block_table_tensor.device,
+                dtype=block_table_tensor.dtype,
+            )[None, None, :]
+        )
+        block_table_tensor = block_table_tensor.view(bs, -1)
+
+        # after remapping, we assume the block size already equals to 1
+
+        max_blk_size_per_req = block_table_tensor.shape[-1]
+        mask = torch.arange(
+            block_table_tensor.size(1), dtype=block_table_tensor.dtype, device=device
+        ).unsqueeze(0) < seq_lens_device.unsqueeze(1)
+        paged_kv_indices = block_table_tensor[mask]
+
+        paged_kv_last_page_len = seq_lens_device % page_size
+        paged_kv_last_page_len = torch.where(
+            paged_kv_last_page_len == 0, page_size, paged_kv_last_page_len
+        )
+
+        paged_kv_indptr = torch.cat(
+            [
+                torch.zeros(1, dtype=seq_lens_device.dtype, device=device),
+                seq_lens_device.cumsum(dim=0, dtype=torch.int32),
+            ]
+        )
+
+        if self.compilation_config.cudagraph_mode.has_full_cudagraphs():
+            num_actual_pages = paged_kv_indices.size(0)
+            self.block_table_remapping[:num_reqs, :max_blk_size_per_req].copy_(
+                block_table_tensor, non_blocking=True
+            )
+            block_table_tensor = self.block_table_remapping[
+                :num_reqs, :max_blk_size_per_req
+            ]
+
+            self.paged_kv_indices[:num_actual_pages].copy_(
+                paged_kv_indices, non_blocking=True
+            )
+            self.paged_kv_indices[num_actual_pages:].fill_(-1)
+            paged_kv_indices = self.paged_kv_indices[:num_actual_pages]
+
+            self.paged_kv_indptr[: 1 + num_reqs].copy_(
+                paged_kv_indptr, non_blocking=True
+            )
+            self.paged_kv_indptr[1 + num_reqs :].fill_(paged_kv_indptr[-1])
+            paged_kv_indptr = self.paged_kv_indptr[: 1 + num_reqs]
+
+            self.paged_kv_last_page_len[:num_reqs].copy_(
+                paged_kv_last_page_len, non_blocking=True
+            )
+            self.paged_kv_last_page_len[num_reqs:].fill_(1)
+            paged_kv_last_page_len = self.paged_kv_last_page_len[:num_reqs]
+
+            qo_indptr = self.qo_indptr[: 1 + num_reqs]
+
+        else:
+            qo_indptr = torch.arange(
+                0, num_reqs + 1, step=1, dtype=torch.int32, device=device
+            )
+
+        attn_metadata = AiterMLADecodeMetadata(
+            block_table=block_table_tensor,
+            seq_lens=seq_lens_device,
+            paged_kv_indptr=paged_kv_indptr,
+            paged_kv_indices=paged_kv_indices,
+            paged_kv_last_page_len=paged_kv_last_page_len,
+            qo_indptr=qo_indptr,
+            dcp_tot_seq_lens=dcp_tot_seq_lens_device,
+        )
+
+        return attn_metadata
+
+
+class AiterMLAImpl(MLACommonImpl[AiterMLAMetadata]):
+    def __init__(
+        self,
+        num_heads: int,
+        head_size: int,
+        scale: float,
+        num_kv_heads: int,
+        alibi_slopes: list[float] | None,
+        sliding_window: int | None,
+        kv_cache_dtype: str,
+        logits_soft_cap: float | None,
+        attn_type: str,
+        kv_sharing_target_layer_name: str | None,
+        # MLA Specific Arguments
+        **mla_args,
+    ) -> None:
+        super().__init__(
+            num_heads,
+            head_size,
+            scale,
+            num_kv_heads,
+            alibi_slopes,
+            sliding_window,
+            kv_cache_dtype,
+            logits_soft_cap,
+            attn_type,
+            kv_sharing_target_layer_name,
+            **mla_args,
+        )
+        assert num_heads == 16 or num_heads == 128, (
+            f"Aiter MLA only supports 16 or 128 number of heads.\n"
+            f"Provided {num_heads} number of heads.\n"
+            "Try adjusting tensor_parallel_size value."
+        )
+        unsupported_features = [alibi_slopes, sliding_window, logits_soft_cap]
+        if any(unsupported_features):
+            raise NotImplementedError(
+                "Aiter MLA does not support one of the following: "
+                "alibi_slopes, sliding_window, logits_soft_cap"
+            )
+
+        from aiter import flash_attn_varlen_func
+
+        self.flash_attn_varlen_func = flash_attn_varlen_func
+
+    def _flash_attn_varlen_diff_headdims(
+        self, q, k, v, return_softmax_lse=False, softmax_scale=None, **kwargs
+    ):
+        output = self.flash_attn_varlen_func(
+            q=q,
+            k=k,
+            v=v,
+            softmax_scale=softmax_scale,
+            return_lse=return_softmax_lse,
+            **kwargs,
+        )
+
+        return output
+
+    def _forward_decode(
+        self,
+        q: torch.Tensor | tuple[torch.Tensor, torch.Tensor],
+        kv_c_and_k_pe_cache: torch.Tensor,
+        attn_metadata: AiterMLAMetadata,
+        layer: AttentionLayer,
+    ) -> tuple[torch.Tensor, torch.Tensor | None]:
+        assert kv_c_and_k_pe_cache.numel() > 0
+        assert attn_metadata.decode is not None
+
+        if type(q) is tuple:
+            q = torch.cat(q, dim=-1)
+
+        assert isinstance(q, torch.Tensor)
+        B = q.shape[0]
+        o = torch.zeros(
+            B, self.num_heads, self.kv_lora_rank, dtype=q.dtype, device=q.device
+        )
+
+        kv_buffer = kv_c_and_k_pe_cache.unsqueeze(2)
+
+        # max_seqlen_qo must be 1 except for MTP
+        # TODO: Find the best value for MTP
+        max_seqlen_qo = 1
+        rocm_aiter_ops.mla_decode_fwd(
+            q,
+            kv_buffer,
+            o,
+            self.scale,
+            attn_metadata.decode.qo_indptr,
+            max_seqlen_qo,
+            attn_metadata.decode.paged_kv_indptr,
+            attn_metadata.decode.paged_kv_indices,
+            attn_metadata.decode.paged_kv_last_page_len,
+        )
+
+        return o, None