Dev gzf deepspeed (#1833)

zhifu gao

2024-06-20 e65b1f701abca03bf3a1b5fbb200392aabd38c22

 funasr/models/paraformer/cif_predictor.py

@@ -494,6 +494,8 @@
        token_num_floor = torch.floor(token_num)



        return hidden, alphas, token_num_floor





@torch.jit.script

def cif_v1_export(hidden, alphas, threshold: float):

    device = hidden.device

@@ -516,9 +518,7 @@
    fires[fire_idxs] = 1

    fires = fires + prefix_sum - prefix_sum_floor



    prefix_sum_hidden = torch.cumsum(

        alphas.unsqueeze(-1).tile((1, 1, hidden_size)) * hidden, dim=1

    )

    prefix_sum_hidden = torch.cumsum(alphas.unsqueeze(-1).tile((1, 1, hidden_size)) * hidden, dim=1)



    frames = prefix_sum_hidden[fire_idxs]

    shift_frames = torch.roll(frames, 1, dims=0)

@@ -530,9 +530,7 @@
    shift_frames[shift_batch_idxs] = 0



    remains = fires - torch.floor(fires)

    remain_frames = (

        remains[fire_idxs].unsqueeze(-1).tile((1, hidden_size)) * hidden[fire_idxs]

    )

    remain_frames = remains[fire_idxs].unsqueeze(-1).tile((1, hidden_size)) * hidden[fire_idxs]



    shift_remain_frames = torch.roll(remain_frames, 1, dims=0)

    shift_remain_frames[shift_batch_idxs] = 0

@@ -541,13 +539,12 @@


    max_label_len = batch_len.max()



    frame_fires = torch.zeros(

        batch_size, max_label_len, hidden_size, dtype=dtype, device=device

    )

    frame_fires = torch.zeros(batch_size, max_label_len, hidden_size, dtype=dtype, device=device)

    indices = torch.arange(max_label_len, device=device).expand(batch_size, -1)

    frame_fires_idxs = indices < batch_len.unsqueeze(1)

    frame_fires[frame_fires_idxs] = frames

    return frame_fires, fires





@torch.jit.script

def cif_export(hidden, alphas, threshold: float):

@@ -661,14 +658,13 @@
    return torch.stack(list_ls, 0), fires





def cif_v1(hidden, alphas, threshold):

def cif_wo_hidden_v1(alphas, threshold, return_fire_idxs=False):

    batch_size, len_time = alphas.size()

    device = alphas.device

    dtype = alphas.dtype



    device = hidden.device

    dtype = hidden.dtype

    batch_size, len_time, hidden_size = hidden.size()

    threshold = torch.tensor([threshold], dtype=alphas.dtype).to(alphas.device)



    frames = torch.zeros(batch_size, len_time, hidden_size, dtype=dtype, device=device)

    fires = torch.zeros(batch_size, len_time, dtype=dtype, device=device)



    prefix_sum = torch.cumsum(alphas, dim=1)

@@ -682,10 +678,19 @@
    fire_idxs = dislocation_diff > 0

    fires[fire_idxs] = 1

    fires = fires + prefix_sum - prefix_sum_floor

    if return_fire_idxs:

        return fires, fire_idxs

    return fires



    prefix_sum_hidden = torch.cumsum(

        alphas.unsqueeze(-1).tile((1, 1, hidden_size)) * hidden, dim=1

    )



def cif_v1(hidden, alphas, threshold):

    fires, fire_idxs = cif_wo_hidden_v1(alphas, threshold, return_fire_idxs=True)



    device = hidden.device

    dtype = hidden.dtype

    batch_size, len_time, hidden_size = hidden.size()

    frames = torch.zeros(batch_size, len_time, hidden_size, dtype=dtype, device=device)

    prefix_sum_hidden = torch.cumsum(alphas.unsqueeze(-1).tile((1, 1, hidden_size)) * hidden, dim=1)



    frames = prefix_sum_hidden[fire_idxs]

    shift_frames = torch.roll(frames, 1, dims=0)

@@ -697,9 +702,7 @@
    shift_frames[shift_batch_idxs] = 0



    remains = fires - torch.floor(fires)

    remain_frames = (

        remains[fire_idxs].unsqueeze(-1).tile((1, hidden_size)) * hidden[fire_idxs]

    )

    remain_frames = remains[fire_idxs].unsqueeze(-1).tile((1, hidden_size)) * hidden[fire_idxs]



    shift_remain_frames = torch.roll(remain_frames, 1, dims=0)

    shift_remain_frames[shift_batch_idxs] = 0

@@ -708,9 +711,7 @@


    max_label_len = batch_len.max()



    frame_fires = torch.zeros(

        batch_size, max_label_len, hidden_size, dtype=dtype, device=device

    )

    frame_fires = torch.zeros(batch_size, max_label_len, hidden_size, dtype=dtype, device=device)

    indices = torch.arange(max_label_len, device=device).expand(batch_size, -1)

    frame_fires_idxs = indices < batch_len.unsqueeze(1)

    frame_fires[frame_fires_idxs] = frames