Skip to content

vllm.models.minimax_m3.amd.sparse_attention_nvfp4

MiniMax M3 block-sparse attention over an NVFP4 K/V cache on ROCm (gfx950).

The cache spec is NVIDIA's, but the bytes inside a page are laid out for the FlyDSL kernels, so pages are not interchangeable with NVIDIA's.

Classes:

Functions:

MiniMaxM3SparseRocmNvfp4Backend

Bases: MiniMaxM3SparseBackend

MiniMax M3 sparse attend over NVFP4 K/V pages (ROCm, gfx950).

Methods:

  • customize_spec –

    Store each head's K and V as two slots of packed data + block scales.

Source code in vllm/models/minimax_m3/amd/sparse_attention_nvfp4.py 
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
class MiniMaxM3SparseRocmNvfp4Backend(MiniMaxM3SparseBackend):
    """MiniMax M3 sparse attend over NVFP4 K/V pages (ROCm, gfx950)."""

    supported_kv_cache_dtypes: ClassVar[list[CacheDType]] = ["nvfp4"]

    @staticmethod
    def get_builder_cls() -> type["MiniMaxM3SparseRocmNvfp4MetadataBuilder"]:
        return MiniMaxM3SparseRocmNvfp4MetadataBuilder

    @classmethod
    def customize_spec(cls, spec: AttentionSpec) -> AttentionSpec:
        """Store each head's K and V as two slots of packed data + block scales."""
        if spec.state_content_bytes is not None or not spec.kv_quant_mode.is_nvfp4:
            return spec
        return replace(
            spec,
            num_head_slots=2 * spec.num_kv_heads,
            state_content_bytes=nvfp4_kv_cache_full_dim(spec.head_size)
            * get_dtype_size(spec.dtype),
        )

    @classmethod
    def supported_kv_cache_layouts(cls) -> tuple[KVCacheLayout, ...] | None:
        # The kernels read the page stride from stride(0) and address pages in
        # 64 bits, so block-outer packing works too (opt in with BLHNC).
        return (KVCacheLayout.LBHNC, KVCacheLayout.BLHNC)

customize_spec(spec) classmethod

Store each head's K and V as two slots of packed data + block scales.

Source code in vllm/models/minimax_m3/amd/sparse_attention_nvfp4.py 
55
56
57
58
59
60
61
62
63
64
65
@classmethod
def customize_spec(cls, spec: AttentionSpec) -> AttentionSpec:
    """Store each head's K and V as two slots of packed data + block scales."""
    if spec.state_content_bytes is not None or not spec.kv_quant_mode.is_nvfp4:
        return spec
    return replace(
        spec,
        num_head_slots=2 * spec.num_kv_heads,
        state_content_bytes=nvfp4_kv_cache_full_dim(spec.head_size)
        * get_dtype_size(spec.dtype),
    )

MiniMaxM3SparseRocmNvfp4Impl

Bases: MiniMaxM3SparseImpl

FlyDSL NVFP4 sparse decode (bf16 compute) and prefill (fp16 compute).

Methods:

Source code in vllm/models/minimax_m3/amd/sparse_attention_nvfp4.py 
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
class MiniMaxM3SparseRocmNvfp4Impl(MiniMaxM3SparseImpl):
    """FlyDSL NVFP4 sparse decode (bf16 compute) and prefill (fp16 compute)."""

    def __init__(self, *args, **kwargs) -> None:
        super().__init__(*args, **kwargs)
        reason = rocm_nvfp4_unsupported_reason(
            self.num_heads,
            self.num_kv_heads,
            self.head_size,
            self.topk_blocks,
            self.block_size,
        )
        if reason is not None:
            raise ValueError(f"MiniMax M3 NVFP4 KV cache on ROCm {reason}")

    def forward(
        self,
        layer: AttentionLayer,
        query: torch.Tensor,
        kv_cache: torch.Tensor,
        output: torch.Tensor,
    ) -> torch.Tensor:
        from vllm.models.minimax_m3.amd.ops.nvfp4_prefill.host import (
            minimax_m3_nvfp4_sparse_prefill,
        )
        from vllm.models.minimax_m3.amd.ops.nvfp4_sparse_decode.host import (
            minimax_m3_nvfp4_sparse_decode_fly,
        )

        attn_metadata = get_forward_context().attn_metadata
        if not isinstance(attn_metadata, dict):
            return output  # profiling run; caches unbound
        md = attn_metadata[layer.layer_name]  # type: ignore[attr-defined]
        assert isinstance(md, MiniMaxM3SparseMetadata)

        nd = md.num_decode_tokens
        num_tokens = md.num_actual_tokens
        topk = layer.topk_indices_buffer  # type: ignore[attr-defined]
        assert topk is not None
        q = query[:num_tokens].view(-1, self.num_heads, self.head_size)
        out = output[:num_tokens].view(-1, self.num_heads, self.head_size)
        k_scale = layer._k_scale  # type: ignore[attr-defined]
        v_scale = layer._v_scale  # type: ignore[attr-defined]

        if md.num_decodes > 0:
            d = md.decode
            assert d is not None
            minimax_m3_nvfp4_sparse_decode_fly(
                q[:nd].contiguous(),
                kv_cache,
                topk[:, :nd, :],
                d.block_table,
                d.seq_lens,
                self.scale,
                out[:nd],
                k_scale=k_scale,
                v_scale=v_scale,
                decode_query_len=d.decode_query_len,
            )
        if md.num_prefills > 0:
            p = md.prefill
            assert p is not None
            assert p.query_req_id is not None and p.query_abs_pos is not None
            minimax_m3_nvfp4_sparse_prefill(
                q[nd:],
                kv_cache,
                topk[:, nd:num_tokens, :],
                p.block_table,
                p.query_req_id,
                p.query_abs_pos,
                self.num_kv_heads,
                self.scale,
                out[nd:],
                k_scale=k_scale,
                v_scale=v_scale,
            )
        return output

    def forward_decode_mxfp8(
        self,
        layer: AttentionLayer,
        query: torch.Tensor,
        kv_cache: torch.Tensor,
        out_q: torch.Tensor,
        out_scale: torch.Tensor,
    ) -> None:
        """``forward`` of a decode-only batch with the output emitted as MXFP8
        (E4M3 values ``out_q`` [tokens, H * D], row-major E8M0 scales
        ``out_scale`` [tokens, H * D / 32]) for the o_proj that follows, instead
        of bf16. Equal to ``mxfp8_e4m3_quantize`` of the bf16 output."""
        from vllm.models.minimax_m3.amd.ops.nvfp4_sparse_decode.host import (
            minimax_m3_nvfp4_sparse_decode_fly,
        )

        md = get_forward_context().attn_metadata[layer.layer_name]  # type: ignore[attr-defined,index]
        assert isinstance(md, MiniMaxM3SparseMetadata)
        assert md.num_prefills == 0 and md.num_decodes > 0
        d = md.decode
        assert d is not None
        nd = md.num_decode_tokens
        topk = layer.topk_indices_buffer  # type: ignore[attr-defined]
        assert topk is not None
        q = query[:nd].view(-1, self.num_heads, self.head_size)
        minimax_m3_nvfp4_sparse_decode_fly(
            q.contiguous(),
            kv_cache,
            topk[:, :nd, :],
            d.block_table,
            d.seq_lens,
            self.scale,
            None,
            k_scale=layer._k_scale,  # type: ignore[attr-defined]
            v_scale=layer._v_scale,  # type: ignore[attr-defined]
            decode_query_len=d.decode_query_len,
            out_q=out_q[:nd],
            out_scale=out_scale[:nd],
        )

forward_decode_mxfp8(layer, query, kv_cache, out_q, out_scale)

forward of a decode-only batch with the output emitted as MXFP8 (E4M3 values out_q [tokens, H * D], row-major E8M0 scales out_scale [tokens, H * D / 32]) for the o_proj that follows, instead of bf16. Equal to mxfp8_e4m3_quantize of the bf16 output.

Source code in vllm/models/minimax_m3/amd/sparse_attention_nvfp4.py 
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
def forward_decode_mxfp8(
    self,
    layer: AttentionLayer,
    query: torch.Tensor,
    kv_cache: torch.Tensor,
    out_q: torch.Tensor,
    out_scale: torch.Tensor,
) -> None:
    """``forward`` of a decode-only batch with the output emitted as MXFP8
    (E4M3 values ``out_q`` [tokens, H * D], row-major E8M0 scales
    ``out_scale`` [tokens, H * D / 32]) for the o_proj that follows, instead
    of bf16. Equal to ``mxfp8_e4m3_quantize`` of the bf16 output."""
    from vllm.models.minimax_m3.amd.ops.nvfp4_sparse_decode.host import (
        minimax_m3_nvfp4_sparse_decode_fly,
    )

    md = get_forward_context().attn_metadata[layer.layer_name]  # type: ignore[attr-defined,index]
    assert isinstance(md, MiniMaxM3SparseMetadata)
    assert md.num_prefills == 0 and md.num_decodes > 0
    d = md.decode
    assert d is not None
    nd = md.num_decode_tokens
    topk = layer.topk_indices_buffer  # type: ignore[attr-defined]
    assert topk is not None
    q = query[:nd].view(-1, self.num_heads, self.head_size)
    minimax_m3_nvfp4_sparse_decode_fly(
        q.contiguous(),
        kv_cache,
        topk[:, :nd, :],
        d.block_table,
        d.seq_lens,
        self.scale,
        None,
        k_scale=layer._k_scale,  # type: ignore[attr-defined]
        v_scale=layer._v_scale,  # type: ignore[attr-defined]
        decode_query_len=d.decode_query_len,
        out_q=out_q[:nd],
        out_scale=out_scale[:nd],
    )

MiniMaxM3SparseRocmNvfp4MetadataBuilder

Bases: MiniMaxM3SparseMetadataBuilder

Adds each prefill token's request and absolute position.

Source code in vllm/models/minimax_m3/amd/sparse_attention_nvfp4.py 
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
class MiniMaxM3SparseRocmNvfp4MetadataBuilder(MiniMaxM3SparseMetadataBuilder):
    """Adds each prefill token's request and absolute position."""

    def build(
        self,
        common_prefix_len: int,
        common_attn_metadata: CommonAttentionMetadata,
        fast_build: bool = False,
    ) -> MiniMaxM3SparseMetadata:
        metadata = super().build(common_prefix_len, common_attn_metadata, fast_build)
        p = metadata.prefill
        if p is not None and p.query_req_id is None:
            p.query_req_id, p.query_abs_pos = minimax_m3_query_token_positions(
                p.cu_seqlens_q, p.context_lens, metadata.num_prefill_tokens
            )
        return metadata

rocm_nvfp4_unsupported_reason(num_heads, num_kv_heads, head_dim, topk_blocks, block_size)

Why the ROCm NVFP4 attend cannot serve this layer, None when it can.

Source code in vllm/models/minimax_m3/amd/sparse_attention_nvfp4.py 
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
def rocm_nvfp4_unsupported_reason(
    num_heads: int, num_kv_heads: int, head_dim: int, topk_blocks: int, block_size: int
) -> str | None:
    """Why the ROCm NVFP4 attend cannot serve this layer, None when it can."""
    from vllm.models.minimax_m3.amd.ops.nvfp4_prefill.host import (
        supports_nvfp4_sparse_prefill,
    )

    if not supports_nvfp4_sparse_prefill(
        num_heads, num_kv_heads, head_dim, topk_blocks, block_size
    ):
        return (
            "needs gfx950, head_dim 128, 16 query heads per KV head and 16 "
            f"selected 128-token blocks (got num_heads={num_heads}, "
            f"num_kv_heads={num_kv_heads}, head_dim={head_dim}, "
            f"topk_blocks={topk_blocks}, block_size={block_size})"
        )
    return None