A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 15, 09:45 UTC
[FEAT][ROCm] Upgrade AITER MLA v1 backend (#18338)
Signed-off-by: vllmellm <vllm.ellm@embeddedllm.com> Co-authored-by: Luka Govedič <ProExpertProg@users.noreply.github.com>
vllmellm committed
May 22, 2025 at 01:34 UTC
94d8ec8d2bcb4ec55e33022b313c7e978edf05e1
2 files changed
+31
-7
docker/Dockerfile.rocm_base
+1
-1
@@ -12,7 +12,7 @@ ARG PYTORCH_REPO="https://github.com/pytorch/pytorch.git"
12
ARG PYTORCH_VISION_REPO="https://github.com/pytorch/vision.git"
13
ARG FA_BRANCH="1a7f4dfa"
14
ARG FA_REPO="https://github.com/Dao-AILab/flash-attention.git"
15
-ARG AITER_BRANCH="5a77249"
15
+ARG AITER_BRANCH="c1debd8"
16
ARG AITER_REPO="https://github.com/ROCm/aiter.git"
17
18
FROM ${BASE_IMAGE} AS base
vllm/v1/attention/backends/mla/rocm_aiter_mla.py
+30
-6
@@ -53,6 +53,8 @@ class AiterMLADecodeMetadata(MLACommonDecodeMetadata):
53
# The number of entries in the last page of each request in
54
# the paged kv cache, shape: [batch_size]
55
paged_kv_last_page_len: Optional[torch.Tensor] = None
56
+ # The query indptr, shape : [num_decode + 1]
57
+ qo_indptr: Optional[torch.Tensor] = None
58
59
60
class AiterMLAMetadata(MLACommonMetadata[AiterMLADecodeMetadata]):
@@ -75,27 +77,33 @@ class AiterMLAMetadataBuilder(MLACommonMetadataBuilder[AiterMLAMetadata]):
77
seq_lens: torch.Tensor) -> tuple[torch.Tensor, ...]:
78
page_size = self.runner.block_size
79
block_table_bounds = (seq_lens + page_size - 1) // page_size
80
+ device = self.runner.device
81
82
mask = (torch.arange(block_table.size(1),
83
dtype=block_table.dtype,
81
- device=block_table.device).unsqueeze(0)
84
+ device=device).unsqueeze(0)
85
< block_table_bounds.unsqueeze(1))
86
paged_kv_indices = block_table[mask]
87
88
paged_kv_indptr = torch.cat([
86
- torch.zeros(1,
87
- dtype=block_table_bounds.dtype,
88
- device=block_table_bounds.device),
89
+ torch.zeros(1, dtype=block_table_bounds.dtype, device=device),
90
block_table_bounds.cumsum(dim=0, dtype=torch.int32)
91
])
92
93
paged_kv_last_page_len = seq_lens % page_size
94
paged_kv_last_page_len = torch.where(paged_kv_last_page_len == 0,
95
page_size, paged_kv_last_page_len)
96
+ qo_indptr = torch.arange(0,
97
+ self._num_decodes + 1,
98
+ step=1,
99
+ dtype=torch.int32,
100
+ device=device)
101
+
102
return (
103
paged_kv_indices,
104
paged_kv_indptr,
105
paged_kv_last_page_len,
106
+ qo_indptr,
107
)
108
109
def _build_decode(self, block_table_tensor: torch.Tensor,
@@ -105,6 +113,7 @@ class AiterMLAMetadataBuilder(MLACommonMetadataBuilder[AiterMLAMetadata]):
113
paged_kv_indices,
114
paged_kv_indptr,
115
paged_last_page_len,
116
+ qo_indptr,
117
) = self._get_paged_kv_tensors(block_table_tensor, seq_lens)
118
119
attn_metadata = AiterMLADecodeMetadata(
@@ -112,7 +121,8 @@ class AiterMLAMetadataBuilder(MLACommonMetadataBuilder[AiterMLAMetadata]):
121
seq_lens=seq_lens,
122
paged_kv_indptr=paged_kv_indptr,
123
paged_kv_indices=paged_kv_indices,
115
- paged_kv_last_page_len=paged_last_page_len)
124
+ paged_kv_last_page_len=paged_last_page_len,
125
+ qo_indptr=qo_indptr)
126
127
return attn_metadata
128
@@ -137,7 +147,10 @@ class AiterMLAImpl(MLACommonImpl[AiterMLAMetadata]):
147
alibi_slopes, sliding_window, kv_cache_dtype,
148
blocksparse_params, logits_soft_cap, attn_type,
149
**mla_args)
140
-
150
+ assert (num_heads == 16 or num_heads == 128), (
151
+ f"Aiter MLA only supports 16 or 128 number of heads.\n"
152
+ f"Provided {num_heads} number of heads.\n"
153
+ "Try adjusting tensor_parallel_size value.")
154
unsupported_features = [
155
alibi_slopes, sliding_window, blocksparse_params, logits_soft_cap
156
]
@@ -189,7 +202,18 @@ class AiterMLAImpl(MLACommonImpl[AiterMLAMetadata]):
202
203
kv_buffer = kv_c_and_k_pe_cache.unsqueeze(2)
204
205
+ if self.num_heads == 16:
206
+ # AITER MLA decode kernel only supports
207
+ # max_seqlen_q=1 when using 16 heads.
208
+ max_seqlen_qo = 1
209
+ else:
210
+ # AITER MLA decode Kernel handles arbitrary
211
+ # max_seqlen_q values when using 128 heads.
212
+ assert attn_metadata.prefill is not None
213
+ max_seqlen_qo = attn_metadata.prefill.max_query_len
214
+
215
aiter_mla_decode_fwd(q, kv_buffer, o, self.scale,
216
+ attn_metadata.decode.qo_indptr, max_seqlen_qo,
217
attn_metadata.decode.paged_kv_indptr,
218
attn_metadata.decode.paged_kv_indices,
219
attn_metadata.decode.paged_kv_last_page_len)