Update PatchedVLLMKVCache for deepseek performance (#194)

Co-authored-by: Linoy Buchnik <[email protected]>

Author: mengniwang95

neural_compressor/torch/algorithms/fp8_quant/_quant_common/helper_modules.py

@@ -1067,7 +1067,11 @@ def forward_measure(self, input, cache, *args, **kwargs):
         return output_cache
 
     def fetch_from_cache(self, cache, blocks, permutations=None):
-        quant_cache = self.quant_input(cache)
+        # TODO: Remove this workaround in next release [SW-221595]
+        if cache.dtype != self.lp_dtype:
+            quant_cache = self.quant_input(cache)
+        else:
+            quant_cache = cache
         if permutations:
             output_cache = self.orig_mod.fetch_from_cache(quant_cache, blocks, permutations)
             for i in range(len(output_cache)):