feat: mixed precision quantization

vllm_rbln/__init__.py

@@ -46,6 +46,7 @@ def register_ops():
         import vllm_rbln.attention.layer  # noqa
         import vllm_rbln.model_executor.layers.fused_moe.layer  # noqa
         import vllm_rbln.model_executor.layers.logits_processor  # noqa
+        import vllm_rbln.model_executor.layers.quantization.kernels.mixed_precision  # noqa
         import vllm_rbln.model_executor.layers.rotary_embedding.base  # noqa
         import vllm_rbln.model_executor.layers.rotary_embedding.deepseek_scaling_rope  # noqa
         import vllm_rbln.model_executor.layers.vocab_parallel_embedding  # noqa

vllm_rbln/model_executor/layers/quantization/kernels/mixed_precision/__init__.py

@@ -0,0 +1,65 @@
+# Copyright 2025 Rebellions Inc. All rights reserved.
+
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at:
+
+#     http://www.apache.org/licenses/LICENSE-2.0
+
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# NOTE: This module should be imported before creating model config,
+# because the function we're patching is imported at that moment.
+
+from typing import Optional
+
+import vllm.envs as envs
+import vllm.model_executor.layers.quantization.kernels.mixed_precision as mp
+from vllm.model_executor.layers.quantization.kernels.mixed_precision.MPLinearKernel import (  # noqa: E501
+    MPLinearKernel, MPLinearLayerConfig)
+
+from vllm_rbln.model_executor.layers.quantization.kernels.mixed_precision.unpacked import (  # noqa: E501
+    RBLNInt8UnpackedLinearKernel)
+
+choose_mp_linear_kernel_original = mp.choose_mp_linear_kernel
+
+_POSSIBLE_KERNELS: list[type[MPLinearKernel]] = [
+    RBLNInt8UnpackedLinearKernel,
+]
+
+
+def choose_mp_linear_kernel_rbln(
+    config: MPLinearLayerConfig,
+    compute_capability: Optional[int] = None,
+) -> type[MPLinearKernel]:
+    from vllm.platforms import current_platform
+
+    if "rbln" not in current_platform.get_device_name().lower():
+        return choose_mp_linear_kernel_original(config, compute_capability)
+
+    failure_reasons = []
+    for kernel in _POSSIBLE_KERNELS:
+        if kernel.__name__ in envs.VLLM_DISABLED_KERNELS:
+            failure_reasons.append(
+                f' {kernel.__name__} disabled by environment variable')
+            continue
+
+        can_implement, failure_reason = kernel.can_implement(config)
+        if can_implement:
+            return kernel
+        else:
+            failure_reasons.append(
+                f' {kernel.__name__} cannot implement due to: {failure_reason}'
+            )
+
+    raise ValueError(
+        "Failed to find a kernel that can implement the "\
+        "WNA16 linear layer. Reasons: \n"
+        + '\n'.join(failure_reasons))
+
+
+mp.choose_mp_linear_kernel = choose_mp_linear_kernel_rbln

vllm_rbln/model_executor/layers/quantization/kernels/mixed_precision/unpacked.py

@@ -0,0 +1,97 @@
+# Copyright 2025 Rebellions Inc. All rights reserved.
+
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at:
+
+#     http://www.apache.org/licenses/LICENSE-2.0
+
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import os
+from typing import Optional
+
+import torch
+from compressed_tensors.compressors.quantized_compressors import (
+    unpack_from_int32)
+from vllm.model_executor.layers.quantization.kernels.mixed_precision.MPLinearKernel import (  # noqa: E501
+    MPLinearKernel, MPLinearLayerConfig)
+from vllm.model_executor.parameter import BasevLLMParameter
+from vllm.scalar_type import scalar_types
+
+
+class RBLNInt8UnpackedLinearKernel(MPLinearKernel):
+    """
+    Torch native implementation of mixed precision Linear, based on
+    compressed_tensors' dequantize() function. rebel_compiler detects this
+    pattern and maps it to the actual kernel.
+    """
+
+    @classmethod
+    def get_min_capability(cls) -> int:
+        raise NotImplementedError
+
+    @classmethod
+    def can_implement(cls,
+                      c: MPLinearLayerConfig) -> tuple[bool, Optional[str]]:
+        if c.weight_type not in (scalar_types.uint4b8, scalar_types.uint8b128):
+            return False, f"Weight type {c.weight_type} not supported"
+        if c.zero_points:
+            return False, "Asymmetric quantization not supported"
+        if c.group_size not in (-1, 64, 128):
+            return False, f"Group size {c.group_size} not supported"
+        if c.has_g_idx:
+            return False, "Group/dynamic activation ordering not supported"
+        return True, None
+
+    def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
+        c = self.config
+        bits = c.weight_type.mantissa
+
+        os.environ['RBLN_QUANT_BITS'] = str(bits)
+
+        if c.has_g_idx:
+            w_gidx = getattr(layer, self.w_gidx_name)
+            layer.perm = torch.argsort(w_gidx)
+
+        def transform_w_q(x: BasevLLMParameter):
+            in_features, out_features = c.full_weight_shape
+            x.data = unpack_from_int32(x.data, bits,
+                                       torch.Size((out_features, in_features)))
+            if c.has_g_idx:
+                x.data = x.data[:, layer.perm]
+            return x
+
+        def transform_w_s(x: BasevLLMParameter):
+            if c.group_size == 128:
+                # Currently we only support group size 64 natively. So
+                # duplicate scale to break a group into two groups of size 64.
+                x.data = x.data.repeat_interleave(2, dim=-1)
+            return x
+
+        self._transform_param(layer, self.w_q_name, transform_w_q)
+        self._transform_param(layer, self.w_s_name, transform_w_s)
+
+    def apply_weights(self,
+                      layer: torch.nn.Module,
+                      x: torch.Tensor,
+                      bias: Optional[torch.Tensor] = None) -> torch.Tensor:
+        in_features, out_features = self.config.full_weight_shape
+
+        if self.config.has_g_idx:
+            x = x[..., layer.perm]
+
+        w_q, w_s, _, _ = self._get_weight_params(layer)
+        if self.config.group_size > 0:
+            w_q = w_q.view(out_features, in_features // 64,
+                           64)  # see transform_w_s
+            w_fp = w_q.type(x.dtype) * w_s.unsqueeze(-1)
+            w_fp = w_fp.view(out_features, in_features)
+        else:
+            w_fp = w_q.type(x.dtype) * w_s
+
+        return torch.nn.functional.linear(x, w_fp, bias)