Skip to content

vllm_gaudi.ops.ssd_combined

_mamba_chunk_scan_combined_fwd 

_mamba_chunk_scan_combined_fwd(
    x,
    dt,
    A,
    B,
    C,
    chunk_size,
    out,
    D=None,
    z=None,
    dt_bias=None,
    initial_states=None,
    cu_seqlens=None,
    last_chunk_indices=None,
    dt_softplus=False,
    dt_limit=(0.0, float("inf")),
    state_dtype=None,
)
Source code in vllm_gaudi/ops/ssd_combined.py 
 22
 23
 24
 25
 26
 27
 28
 29
 30
 31
 32
 33
 34
 35
 36
 37
 38
 39
 40
 41
 42
 43
 44
 45
 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
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 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
def _mamba_chunk_scan_combined_fwd(
        x,
        dt,
        A,
        B,
        C,
        chunk_size,
        out,
        D=None,
        z=None,
        dt_bias=None,
        initial_states=None,
        cu_seqlens=None,
        last_chunk_indices=None,
        dt_softplus=False,
        dt_limit=(0.0, float("inf")),
        state_dtype=None,
):
    assert is_int_pow_2(chunk_size), "chunk_size must be integer power of 2"
    seqlen, nheads, headdim = x.shape
    _, ngroups, dstate = B.shape
    assert nheads % ngroups == 0
    assert B.shape == (seqlen, ngroups, dstate)
    assert dt.shape == (seqlen, nheads)
    assert A.shape == (nheads, )
    assert C.shape == B.shape
    if z is not None:
        assert z.shape == x.shape
    if D is not None:
        assert D.shape == (nheads, headdim) or D.shape == (nheads, )
    if B.stride(-1) != 1:
        B = B.contiguous()
    if C.stride(-1) != 1:
        C = C.contiguous()
    if (x.stride(-1) != 1 and x.stride(0) != 1):  # Either M or K dimension should be contiguous
        x = x.contiguous()
    if (z is not None and z.stride(-1) != 1 and z.stride(0) != 1):  # Either M or K dimension should be contiguous
        z = z.contiguous()
    if D is not None and D.stride(-1) != 1:
        D = D.contiguous()
    assert cu_seqlens is not None, "Assuming varlen input - must supply cu_seqlens"

    if initial_states is not None:
        assert initial_states.shape == (len(cu_seqlens) - 1, nheads, headdim, dstate)

    # This function executes 5 sub-functions for computing mamba
    # - a good resource is the blog https://goombalab.github.io/blog/2024/mamba2-part3-algorithm/
    #   which has a minimal implementation to understand the below operations
    # - as explained by the blog, mamba is a special case of causal attention
    # - the idea is to chunk the attention matrix and compute each
    #   submatrix separately using different optimizations.
    # - see the blog and paper for a visualization of the submatrices
    #   which we refer to in the comments below

    # 1. Compute chunked cumsum of A * dt
    # - here dt may go through a softplus activation
    dA_cumsum, dt = new_chunk_cumsum(
        dt,
        A,
        chunk_size,
        dt_bias=dt_bias,
        dt_softplus=dt_softplus,
        dt_limit=dt_limit,
    )

    # 2. Compute the state for each intra-chunk
    # (right term of low-rank factorization of off-diagonal blocks; B terms)
    states = new_chunk_state(B, x, dt, dA_cumsum, states_in_fp32=True)

    # 3. Compute the inter-chunk SSM recurrence; produces correct SSM states at chunk boundaries
    # (middle term of factorization of off-diag blocks; A terms)
    # - for handling chunked prefill, this requires initial_states
    states = new_ssd_state_passing(
        rearrange(states, "... p n -> ... (p n)"),
        dA_cumsum,  # (nheads, nchunks, chunk_size)
        initial_states=rearrange(initial_states, "... p n -> ... (p n)")
        if initial_states is not None else None,  # (batch, nheads, headdim*dstate)
        out_dtype=state_dtype if state_dtype is not None else C.dtype,
    )
    states = rearrange(states, "... (p n) -> ... p n", n=dstate)

    # 4. Compute batched matrix multiply for C_j^T B_i terms
    CB = new_ssd_bmm(C, B, chunk_size, output_dtype=torch.float32)

    # 5. Scan and compute the diagonal blocks, taking into
    #    account past causal states.
    # - if initial states are provided, then states information will be
    #   augmented with initial_states.
    # - to do this properly, we need to account for example changes in
    #   the continuous batch, therefore we introduce pseudo chunks, which is
    #   a chunk that is split up each time an example changes.
    new_chunk_scan(
        CB,
        x,
        dt,
        dA_cumsum,
        C,
        states,
        out,  # in-place update
        D=D,
        z=z,
        initial_states=initial_states,
    )

    return states

hpu_mamba_chunk_scan_combined_varlen 

hpu_mamba_chunk_scan_combined_varlen(
    x,
    dt,
    A,
    B,
    C,
    chunk_size,
    cu_seqlens,
    last_chunk_indices,
    out,
    D=None,
    z=None,
    dt_bias=None,
    initial_states=None,
    dt_softplus=False,
    dt_limit=(0.0, float("inf")),
    state_dtype=None,
)
Argument

x: (seqlen, nheads, headdim) dt: (seqlen, nheads) A: (nheads) B: (seqlen, ngroups, dstate) C: (seqlen, ngroups, dstate) chunk_size: int cu_seqlens: (batch + 1,) last_chunk_indices: (batch,) out: (seqlen, nheads, headdim) preallocated output tensor D: (nheads, headdim) or (nheads,) z: (seqlen, nheads, headdim) dt_bias: (nheads,) initial_states: (batch, nheads, headdim, dstate) dt_softplus: Whether to apply softplus to dt out: (seqlen, nheads, headdim) preallocated output tensor state_dtype: The data type of the ssm state

Return: varlen_states: (batch, nheads, headdim, dstate)

Source code in vllm_gaudi/ops/ssd_combined.py 
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
def hpu_mamba_chunk_scan_combined_varlen(
        x,
        dt,
        A,
        B,
        C,
        chunk_size,
        cu_seqlens,
        last_chunk_indices,
        out,
        D=None,
        z=None,
        dt_bias=None,
        initial_states=None,
        dt_softplus=False,
        dt_limit=(0.0, float("inf")),
        state_dtype=None,
):
    """
    Argument:
        x: (seqlen, nheads, headdim)
        dt: (seqlen, nheads)
        A: (nheads)
        B: (seqlen, ngroups, dstate)
        C: (seqlen, ngroups, dstate)
        chunk_size: int
        cu_seqlens: (batch + 1,)
        last_chunk_indices: (batch,)
        out: (seqlen, nheads, headdim) preallocated output tensor
        D: (nheads, headdim) or (nheads,)
        z: (seqlen, nheads, headdim)
        dt_bias: (nheads,)
        initial_states: (batch, nheads, headdim, dstate)
        dt_softplus: Whether to apply softplus to dt
        out: (seqlen, nheads, headdim) preallocated output tensor
        state_dtype: The data type of the ssm state
    Return:
        varlen_states: (batch, nheads, headdim, dstate)
    """

    assert cu_seqlens is not None, "cu_seqlens must be provided assuming varlen input"

    varlen_states = _mamba_chunk_scan_combined_fwd(
        x,
        dt,
        A,
        B,
        C,
        chunk_size,
        out,
        D=D,
        z=z,
        dt_bias=dt_bias,
        initial_states=initial_states,
        cu_seqlens=cu_seqlens,
        last_chunk_indices=last_chunk_indices,
        dt_softplus=dt_softplus,
        dt_limit=dt_limit,
        state_dtype=state_dtype,
    )

    return varlen_states

is_int_pow_2 

is_int_pow_2(n)
Source code in vllm_gaudi/ops/ssd_combined.py 
18
19
def is_int_pow_2(n):
    return isinstance(n, int) and n > 0 and (n & (n - 1)) == 0