diff --git a/requirements.txt b/requirements.txt
index 3e71feab0..b05282668 100644
--- a/requirements.txt
+++ b/requirements.txt
@@ -4,3 +4,4 @@ pandas>=2.2.3
 numba>=0.58.0
 numpy>=1.26.0
 transformers>=4.1,<4.56.0
+neural-compressor-pt @ file:///software/users/dsemiat/dynamic_quant/vllm/neural_compressor_pt-3.5.dev4+ga3204838c-py3-none-any.whl
diff --git a/vllm_gaudi/extension/ops.py b/vllm_gaudi/extension/ops.py
index 4223a4219..078b01451 100644
--- a/vllm_gaudi/extension/ops.py
+++ b/vllm_gaudi/extension/ops.py
@@ -54,12 +54,15 @@ def block2batch(tensor, block_mapping, matmul_op=torch.matmul):
 
 def pipelined_pa(attn, value, block_bias, block_groups, block_mapping, batch_size, matmul_av_op, batch2block_matmul_op,
                  block2batch_matmul_op):
+    # print(f"Danny running pipelined_pa {get_config().fused_block_softmax=} {get_config().fused_block_softmax_adjustment=}")
     # When fp32_softmax is enabled attn is left in fp32 after Q@K
     # We can return to native dtype after we renormalize and calculate the adjustments
     if block_bias is not None and attn.dtype != block_bias.dtype:
         block_bias = block_bias.to(dtype=attn.dtype)
     # TODO: w/a with 5D req as the block_softmax kernel does not support 4D attn tensor, which is used in e.g. Granite-3B
     if get_config().fused_block_softmax and get_config().fused_block_softmax_adjustment and attn.dim() == 5:
+        # print("Danny Warning: fused_block_softmax with 5D attn tensor is not supported yet, falling back to unfused path.")
+        # print("Danny using fused_block_softmax without calling the nn.Module member.")
         attn, block_max, block_sums = torch.ops.hpu.block_softmax(attn, block_bias, block_groups)
         if attn.dtype == torch.float32:
             attn = attn.to(value.dtype)
@@ -74,6 +77,7 @@ def pipelined_pa(attn, value, block_bias, block_groups, block_mapping, batch_siz
         block_sums = attn.sum(dim=-1, keepdim=True)
     attn = matmul_av_op(attn, value)
     if get_config().fused_block_softmax_adjustment:
+        # print("Danny using block_softmax_adjustment without calling the nn.Module member.")
         out_shape = list(attn.shape[:3]) + [1] * (attn.dim() - 3)
         rescale = torch.ops.hpu.block_softmax_adjustment(block_max, block_sums.to(block_max.dtype), block_groups,
                                                          batch_size, out_shape).to(attn.dtype)
@@ -476,6 +480,7 @@ def forward(self, hidden_states, expert_routing_table, router_weights, permuted_
         w2_list = [self.w2_list[i].weight.squeeze() for i in experts_range]
 
         if self.moe_n_slice == 1:
+            print(f"VllmMixtureOfExpertsOp: running with single slice, calling moe with {hidden_states.shape=}, {expert_routing_table.shape=}, {router_weights.shape=}, {activation=}")
             return torch.ops.hpu.mixture_of_experts(hidden_states=hidden_states,
                                                     expert_routing_table=expert_routing_table,
                                                     router_weights=router_weights,
diff --git a/vllm_gaudi/extension/utils.py b/vllm_gaudi/extension/utils.py
index e5630c0ff..895503cf0 100644
--- a/vllm_gaudi/extension/utils.py
+++ b/vllm_gaudi/extension/utils.py
@@ -26,8 +26,8 @@ class Matmul(torch.nn.Module):
     def __init__(self):
         super(Matmul, self).__init__()
 
-    def forward(self, x, y):
-        return torch.matmul(x, y)
+    def forward(self, x, y, *args, **kwargs):
+        return torch.matmul(x, y, *args, **kwargs)
 
 
 class Softmax(torch.nn.Module):