diff --git a/requirements-hpu.txt b/requirements-hpu.txt
index 07f9c31117e49..ddf1caccf41d8 100644
--- a/requirements-hpu.txt
+++ b/requirements-hpu.txt
@@ -8,7 +8,5 @@ pandas
 tabulate
 setuptools>=61
 setuptools-scm>=8
-vllm-hpu-extension @ git+https://github.com/HabanaAI/vllm-hpu-extension.git@61334c5
+vllm-hpu-extension @ git+https://github.com/HabanaAI/vllm-hpu-extension.git@ac9740d
 neural-compressor @ git+https://github.com/intel/neural-compressor.git@b196432
-
-
diff --git a/vllm/worker/hpu_worker.py b/vllm/worker/hpu_worker.py
index 2b8f955265792..1004af0eca40a 100644
--- a/vllm/worker/hpu_worker.py
+++ b/vllm/worker/hpu_worker.py
@@ -166,7 +166,9 @@ def determine_num_available_blocks(self) -> Tuple[int, int]:
         if is_fake_hpu():
             cache_block_size = self.get_cache_block_size_bytes()
             fake_hpu_cache_alloc = 4 * 2**30  # take 4 GiB flat on fake hpu
-            return fake_hpu_cache_alloc // cache_block_size, 0
+            num_fake_hpu_blocks = fake_hpu_cache_alloc // cache_block_size
+            self.model_runner.bucketing_ctx.num_hpu_blocks = num_fake_hpu_blocks
+            return num_fake_hpu_blocks, 0
         with HabanaMemoryProfiler() as m:
             self.model_runner.profile_run()
             torch.hpu.synchronize()
@@ -203,6 +205,8 @@ def determine_num_available_blocks(self) -> Tuple[int, int]:
         num_hpu_blocks = max(num_hpu_blocks, 0)
         num_cpu_blocks = max(num_cpu_blocks, 0)
 
+        self.model_runner.bucketing_ctx.num_hpu_blocks = num_hpu_blocks
+
         if self.model_runner.lora_manager:
             self.model_runner.remove_all_loras()