Skip to content

vllm_gaudi.ops.hpu_row_parallel_linear

_registered module-attribute 

_registered = False

HPURowParallelLinear

Bases: RowParallelLinear

HPU-optimized RowParallelLinear implementation.

This implementation provides chunked computation for overlapping compute and communication on HPU devices.

Source code in vllm_gaudi/ops/hpu_row_parallel_linear.py 
 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
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
211
212
213
214
215
class HPURowParallelLinear(RowParallelLinear):
    """HPU-optimized RowParallelLinear implementation.

    This implementation provides chunked computation for overlapping
    compute and communication on HPU devices.
    """

    def __init__(self, *args, **kwargs):
        """Initialize HPURowParallelLinear with chunking support.

        The number of chunks can be configured via the row_parallel_chunks
        feature flag (env var VLLM_ROW_PARALLEL_CHUNKS). Default is 1 (disabled).

        The token threshold for enabling chunking can be configured via the
        row_parallel_chunk_threshold feature flag (env var
        VLLM_ROW_PARALLEL_CHUNK_THRESHOLD). Default is 8192 tokens.
        """
        super().__init__(*args, **kwargs)
        config = get_config()
        self.num_chunks = config.row_parallel_chunks
        if self.num_chunks < 1:
            raise ValueError(f"row_parallel_chunks must be >= 1, got {self.num_chunks}")

        self.chunk_threshold = config.row_parallel_chunk_threshold
        if self.chunk_threshold < 1:
            raise ValueError(f"row_parallel_chunk_threshold must be >= 1, got {self.chunk_threshold}")

    def forward(
        self,
        input_,
    ) -> Union[torch.Tensor, tuple[torch.Tensor, Parameter]]:
        """Forward pass with HPU-specific optimizations.

        Args:
            input_: Input tensor to process

        Returns:
            Output tensor, or tuple of (output, bias) if skip_bias_add is True
        """

        if self.input_is_parallel:
            input_parallel = input_
        else:
            splitted_input = split_tensor_along_last_dim(input_, num_partitions=self.tp_size)
            input_parallel = splitted_input[self.tp_rank].contiguous()

        assert self.quant_method is not None

        # Determine total tokens for chunking decision
        input_shape = input_parallel.shape
        # For 3D input [batch, seq, hidden], total_tokens = batch * seq
        # For 2D input [tokens, hidden], total_tokens = tokens
        if input_parallel.ndim == 3:
            batch_size, seq_len, _ = input_parallel.shape
            total_tokens = batch_size * seq_len
        else:
            total_tokens = input_shape[0]

        # Check if we should use chunking
        # Don't chunk for inputs below threshold as there's no overlap benefit
        should_chunk = (self.num_chunks > 1 and self.reduce_results and self.tp_size > 1
                        and total_tokens >= self.chunk_threshold)

        # Chunked computation for overlapping compute and communication
        if should_chunk:
            torch._dynamo.graph_break()

            # Determine if input is 3D [batch, seq, hidden] or 2D [batch*seq, hidden]
            is_3d = input_parallel.ndim == 3

            if is_3d:
                # Input is [batch, seq, hidden]
                # For multi-token sequences (prompts): chunk along sequence dimension
                # For single-token batches (decodes): chunk along batch dimension
                batch_size, seq_len, hidden_dim = input_parallel.shape
                if seq_len > 1:
                    # Chunk along sequence dimension for prompts
                    chunk_dim = 1
                    total_tokens = seq_len
                else:
                    # Chunk along batch dimension for batched decodes
                    chunk_dim = 0
                    total_tokens = batch_size
                output = torch.empty(batch_size,
                                     seq_len,
                                     self.output_size_per_partition,
                                     dtype=input_parallel.dtype,
                                     device=input_parallel.device)
            else:
                # Input is [batch*seq, hidden], chunk along batch dimension
                total_tokens, hidden_dim = input_parallel.shape
                chunk_dim = 0
                output = torch.empty(total_tokens,
                                     self.output_size_per_partition,
                                     dtype=input_parallel.dtype,
                                     device=input_parallel.device)

            chunk_size = (total_tokens + self.num_chunks - 1) // self.num_chunks

            # Lists to store chunks and handles
            output_chunks = []
            handles = []
            chunk_ranges = []

            # Phase 1: Compute all chunks and start all-reduces
            for i in range(self.num_chunks):
                torch._dynamo.graph_break()
                start_idx = i * chunk_size
                end_idx = min((i + 1) * chunk_size, total_tokens)

                # Skip empty chunks
                if start_idx >= end_idx:
                    continue

                # Slice input along the appropriate dimension
                if is_3d:
                    if chunk_dim == 1:
                        # Chunking along sequence dimension
                        input_chunk = input_parallel[:, start_idx:end_idx, :]
                    else:
                        # Chunking along batch dimension
                        input_chunk = input_parallel[start_idx:end_idx, :, :]
                else:
                    input_chunk = input_parallel[start_idx:end_idx]

                # Compute on chunk (no bias - will be added after all-reduce)
                output_chunk = self.quant_method.apply(self, input_chunk, bias=None)
                torch._dynamo.graph_break()

                # Start async all-reduce for this chunk
                handle = torch.distributed.all_reduce(output_chunk, group=get_tp_group().device_group, async_op=True)

                # Store chunk, handle, and range info
                output_chunks.append(output_chunk)
                handles.append(handle)
                chunk_ranges.append((start_idx, end_idx))

            # Phase 2: Wait for all handles and combine outputs
            for handle in handles:
                handle.wait()

            torch._dynamo.graph_break()

            # Copy all chunks to output
            for chunk, (start_idx, end_idx) in zip(output_chunks, chunk_ranges):
                if is_3d:
                    if chunk_dim == 1:
                        # Chunked along sequence dimension
                        output[:, start_idx:end_idx, :] = chunk
                    else:
                        # Chunked along batch dimension
                        output[start_idx:end_idx, :, :] = chunk
                else:
                    output[start_idx:end_idx] = chunk

            torch._dynamo.graph_break()

            # Apply bias after all-reduce (only on rank 0 if not skip_bias_add)
            if self.bias is not None and self.tp_rank == 0 and not self.skip_bias_add:
                output = output + self.bias
        else:
            # Original single-shot computation
            bias_ = None if (self.tp_rank > 0 or self.skip_bias_add) else self.bias
            output_parallel = self.quant_method.apply(self, input_parallel, bias_)

            if self.reduce_results and self.tp_size > 1:
                output = tensor_model_parallel_all_reduce(output_parallel)
            else:
                output = output_parallel

        output_bias = self.bias if self.skip_bias_add else None

        if not self.return_bias:
            return output
        return output, output_bias

chunk_threshold instance-attribute 

chunk_threshold = row_parallel_chunk_threshold

num_chunks instance-attribute 

num_chunks = row_parallel_chunks

__init__ 

__init__(*args, **kwargs)

Initialize HPURowParallelLinear with chunking support.

The number of chunks can be configured via the row_parallel_chunks feature flag (env var VLLM_ROW_PARALLEL_CHUNKS). Default is 1 (disabled).

The token threshold for enabling chunking can be configured via the row_parallel_chunk_threshold feature flag (env var VLLM_ROW_PARALLEL_CHUNK_THRESHOLD). Default is 8192 tokens.

Source code in vllm_gaudi/ops/hpu_row_parallel_linear.py 
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
def __init__(self, *args, **kwargs):
    """Initialize HPURowParallelLinear with chunking support.

    The number of chunks can be configured via the row_parallel_chunks
    feature flag (env var VLLM_ROW_PARALLEL_CHUNKS). Default is 1 (disabled).

    The token threshold for enabling chunking can be configured via the
    row_parallel_chunk_threshold feature flag (env var
    VLLM_ROW_PARALLEL_CHUNK_THRESHOLD). Default is 8192 tokens.
    """
    super().__init__(*args, **kwargs)
    config = get_config()
    self.num_chunks = config.row_parallel_chunks
    if self.num_chunks < 1:
        raise ValueError(f"row_parallel_chunks must be >= 1, got {self.num_chunks}")

    self.chunk_threshold = config.row_parallel_chunk_threshold
    if self.chunk_threshold < 1:
        raise ValueError(f"row_parallel_chunk_threshold must be >= 1, got {self.chunk_threshold}")

forward 

forward(input_) -> Union[Tensor, tuple[Tensor, Parameter]]

Forward pass with HPU-specific optimizations.

Parameters:

Name Type Description Default
input_

Input tensor to process

 required

Returns:

Type Description
Union[Tensor, tuple[Tensor, Parameter]]

Output tensor, or tuple of (output, bias) if skip_bias_add is True
Source code in vllm_gaudi/ops/hpu_row_parallel_linear.py 
 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
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
211
212
213
214
215
def forward(
    self,
    input_,
) -> Union[torch.Tensor, tuple[torch.Tensor, Parameter]]:
    """Forward pass with HPU-specific optimizations.

    Args:
        input_: Input tensor to process

    Returns:
        Output tensor, or tuple of (output, bias) if skip_bias_add is True
    """

    if self.input_is_parallel:
        input_parallel = input_
    else:
        splitted_input = split_tensor_along_last_dim(input_, num_partitions=self.tp_size)
        input_parallel = splitted_input[self.tp_rank].contiguous()

    assert self.quant_method is not None

    # Determine total tokens for chunking decision
    input_shape = input_parallel.shape
    # For 3D input [batch, seq, hidden], total_tokens = batch * seq
    # For 2D input [tokens, hidden], total_tokens = tokens
    if input_parallel.ndim == 3:
        batch_size, seq_len, _ = input_parallel.shape
        total_tokens = batch_size * seq_len
    else:
        total_tokens = input_shape[0]

    # Check if we should use chunking
    # Don't chunk for inputs below threshold as there's no overlap benefit
    should_chunk = (self.num_chunks > 1 and self.reduce_results and self.tp_size > 1
                    and total_tokens >= self.chunk_threshold)

    # Chunked computation for overlapping compute and communication
    if should_chunk:
        torch._dynamo.graph_break()

        # Determine if input is 3D [batch, seq, hidden] or 2D [batch*seq, hidden]
        is_3d = input_parallel.ndim == 3

        if is_3d:
            # Input is [batch, seq, hidden]
            # For multi-token sequences (prompts): chunk along sequence dimension
            # For single-token batches (decodes): chunk along batch dimension
            batch_size, seq_len, hidden_dim = input_parallel.shape
            if seq_len > 1:
                # Chunk along sequence dimension for prompts
                chunk_dim = 1
                total_tokens = seq_len
            else:
                # Chunk along batch dimension for batched decodes
                chunk_dim = 0
                total_tokens = batch_size
            output = torch.empty(batch_size,
                                 seq_len,
                                 self.output_size_per_partition,
                                 dtype=input_parallel.dtype,
                                 device=input_parallel.device)
        else:
            # Input is [batch*seq, hidden], chunk along batch dimension
            total_tokens, hidden_dim = input_parallel.shape
            chunk_dim = 0
            output = torch.empty(total_tokens,
                                 self.output_size_per_partition,
                                 dtype=input_parallel.dtype,
                                 device=input_parallel.device)

        chunk_size = (total_tokens + self.num_chunks - 1) // self.num_chunks

        # Lists to store chunks and handles
        output_chunks = []
        handles = []
        chunk_ranges = []

        # Phase 1: Compute all chunks and start all-reduces
        for i in range(self.num_chunks):
            torch._dynamo.graph_break()
            start_idx = i * chunk_size
            end_idx = min((i + 1) * chunk_size, total_tokens)

            # Skip empty chunks
            if start_idx >= end_idx:
                continue

            # Slice input along the appropriate dimension
            if is_3d:
                if chunk_dim == 1:
                    # Chunking along sequence dimension
                    input_chunk = input_parallel[:, start_idx:end_idx, :]
                else:
                    # Chunking along batch dimension
                    input_chunk = input_parallel[start_idx:end_idx, :, :]
            else:
                input_chunk = input_parallel[start_idx:end_idx]

            # Compute on chunk (no bias - will be added after all-reduce)
            output_chunk = self.quant_method.apply(self, input_chunk, bias=None)
            torch._dynamo.graph_break()

            # Start async all-reduce for this chunk
            handle = torch.distributed.all_reduce(output_chunk, group=get_tp_group().device_group, async_op=True)

            # Store chunk, handle, and range info
            output_chunks.append(output_chunk)
            handles.append(handle)
            chunk_ranges.append((start_idx, end_idx))

        # Phase 2: Wait for all handles and combine outputs
        for handle in handles:
            handle.wait()

        torch._dynamo.graph_break()

        # Copy all chunks to output
        for chunk, (start_idx, end_idx) in zip(output_chunks, chunk_ranges):
            if is_3d:
                if chunk_dim == 1:
                    # Chunked along sequence dimension
                    output[:, start_idx:end_idx, :] = chunk
                else:
                    # Chunked along batch dimension
                    output[start_idx:end_idx, :, :] = chunk
            else:
                output[start_idx:end_idx] = chunk

        torch._dynamo.graph_break()

        # Apply bias after all-reduce (only on rank 0 if not skip_bias_add)
        if self.bias is not None and self.tp_rank == 0 and not self.skip_bias_add:
            output = output + self.bias
    else:
        # Original single-shot computation
        bias_ = None if (self.tp_rank > 0 or self.skip_bias_add) else self.bias
        output_parallel = self.quant_method.apply(self, input_parallel, bias_)

        if self.reduce_results and self.tp_size > 1:
            output = tensor_model_parallel_all_reduce(output_parallel)
        else:
            output = output_parallel

    output_bias = self.bias if self.skip_bias_add else None

    if not self.return_bias:
        return output
    return output, output_bias

register 

register()

Conditionally register HPURowParallelLinear as OOT replacement.

Only registers when VLLM_ROW_PARALLEL_CHUNKS > 1, avoiding memory overhead from unconditional layer replacement (GAUDISW-247164).

Source code in vllm_gaudi/ops/hpu_row_parallel_linear.py 
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
def register():
    """Conditionally register HPURowParallelLinear as OOT replacement.

    Only registers when VLLM_ROW_PARALLEL_CHUNKS > 1, avoiding memory
    overhead from unconditional layer replacement (GAUDISW-247164).
    """
    global _registered
    if _registered:
        return
    _registered = True

    env_value = os.environ.get('VLLM_ROW_PARALLEL_CHUNKS', '1')
    try:
        num_chunks = int(env_value)
    except ValueError:
        num_chunks = 1
    if num_chunks > 1:
        RowParallelLinear.register_oot(HPURowParallelLinear)