Port: add VLLM_DISABLE_MARK_SCALES_AS_CONST (#522)

add VLLM_DISABLE_MARK_SCALES_AS_CONST  to avoid too much graph compile.
V0 code base
:https://github.com/HabanaAI/vllm-fork/blob/habana_main/vllm/worker/hpu_model_runner.py#L1228-L1230

Signed-off-by: zhenzhao <zhenzhao@habana.ai>
Co-authored-by: Michał Kuligowski <michal.kuligowski@intel.com>

Author: zhejiangxiaomai

vllm_gaudi/v1/worker/hpu_model_runner.py

@@ -3429,6 +3429,7 @@ def load_model(self) -> None:
             with HabanaMemoryProfiler() as m_inc:
                 from neural_compressor.torch.quantization import (FP8Config, convert, prepare)
                 config = FP8Config.from_json_file(os.getenv("QUANT_CONFIG", ""))
+                disable_mark_scales_as_const = os.getenv("VLLM_DISABLE_MARK_SCALES_AS_CONST", "false") in ("1", "true")
                 self._inc_preprocess()
                 if config.measure:
                     self.model = prepare(self.model, config)
@@ -3437,7 +3438,8 @@ def load_model(self) -> None:
                 else:
                     raise ValueError("Unknown quantization config mode,"
                                      "please validate quantization config file")
-                htcore.hpu_initialize(self.model, mark_only_scales_as_const=True)
+                if not disable_mark_scales_as_const:
+                    htcore.hpu_initialize(self.model, mark_only_scales_as_const=True)
             self.inc_initialized_successfully = True
             self.model_memory_usage = m_inc.consumed_device_memory
             logger.info("Preparing model with INC took %.4f GB", self.model_memory_usage / float(2**30))