Add support to ModelOpt quantization of Mixtral8x7b model

Signed-off-by: Yue <yueshen@nvidia.com>

Author: yueshen2016

vllm/model_executor/models/mixtral_quant.py

@@ -45,7 +45,8 @@
 from vllm.model_executor.layers.sampler import SamplerOutput, get_sampler
 from vllm.model_executor.layers.vocab_parallel_embedding import (
     ParallelLMHead, VocabParallelEmbedding)
-from vllm.model_executor.model_loader.weight_utils import default_weight_loader
+from vllm.model_executor.model_loader.weight_utils import (
+    default_weight_loader, maybe_remap_kv_scale_name)
 from vllm.model_executor.sampling_metadata import SamplingMetadata
 from vllm.sequence import IntermediateTensors
 
@@ -420,6 +421,11 @@ def load_weights(self, weights: Iterable[Tuple[str,
         for name, loaded_weight in weights:
             if "rotary_emb.inv_freq" in name:
                 continue
+            if name.endswith("scale"):
+                # Remapping the name of FP8 kv-scale.
+                name = maybe_remap_kv_scale_name(name, params_dict)
+                if name is None:
+                    continue
             for (param_name, weight_name, shard_id) in stacked_params_mapping:
                 if weight_name not in name:
                     continue