diff --git a/src/mcore_bridge/model/gpts/qwen3_next.py b/src/mcore_bridge/model/gpts/qwen3_next.py
index 6fcf1ea..51a3a5c 100644
--- a/src/mcore_bridge/model/gpts/qwen3_next.py
+++ b/src/mcore_bridge/model/gpts/qwen3_next.py
@@ -470,12 +470,12 @@ def forward(self, hidden_states: torch.Tensor, **kwargs):
         # Note: for packed inputs, we do not perform padding_free unpadding.
         # Doing so would allow different sequences to see each other; for efficiency we keep this implementation.
         if thd_format:
-            max_seqlen_q = int(packed_seq_params.max_seqlen_q)
-            new_hidden_states = hidden_states.new_zeros(
-                (packed_seq_params.num_samples, max_seqlen_q, hidden_states.shape[-1]))
-            attention_mask = hidden_states.new_zeros((packed_seq_params.num_samples, max_seqlen_q), dtype=torch.bool)
             cu_seqlens_q = packed_seq_params.cu_seqlens_q
-            for i in range(packed_seq_params.num_samples):
+            num_samples = cu_seqlens_q.numel() - 1
+            max_seqlen_q = int(packed_seq_params.max_seqlen_q)
+            new_hidden_states = hidden_states.new_zeros((num_samples, max_seqlen_q, hidden_states.shape[-1]))
+            attention_mask = hidden_states.new_zeros((num_samples, max_seqlen_q), dtype=torch.bool)
+            for i in range(num_samples):
                 start, end = cu_seqlens_q[i], cu_seqlens_q[i + 1]
                 attention_mask[i, :end - start] = True
                 new_hidden_states[i, :end - start] = hidden_states[start:end, 0]
diff --git a/src/mcore_bridge/model/mm_gpts/qwen3_5.py b/src/mcore_bridge/model/mm_gpts/qwen3_5.py
index b673bf7..f37ff38 100644
--- a/src/mcore_bridge/model/mm_gpts/qwen3_5.py
+++ b/src/mcore_bridge/model/mm_gpts/qwen3_5.py
@@ -40,12 +40,12 @@ def forward(self, hidden_states: torch.Tensor, **kwargs):
         # Note: for packed inputs, we do not perform padding_free unpadding.
         # Doing so would allow different sequences to see each other; for efficiency we keep this implementation.
         if thd_format:
-            max_seqlen_q = int(packed_seq_params.max_seqlen_q)
-            new_hidden_states = hidden_states.new_zeros(
-                (packed_seq_params.num_samples, max_seqlen_q, hidden_states.shape[-1]))
-            attention_mask = hidden_states.new_zeros((packed_seq_params.num_samples, max_seqlen_q), dtype=torch.bool)
             cu_seqlens_q = packed_seq_params.cu_seqlens_q
-            for i in range(packed_seq_params.num_samples):
+            num_samples = cu_seqlens_q.numel() - 1
+            max_seqlen_q = int(packed_seq_params.max_seqlen_q)
+            new_hidden_states = hidden_states.new_zeros((num_samples, max_seqlen_q, hidden_states.shape[-1]))
+            attention_mask = hidden_states.new_zeros((num_samples, max_seqlen_q), dtype=torch.bool)
+            for i in range(num_samples):
                 start, end = cu_seqlens_q[i], cu_seqlens_q[i + 1]
                 attention_mask[i, :end - start] = True
                 new_hidden_states[i, :end - start] = hidden_states[start:end, 0]
diff --git a/src/mcore_bridge/model/modules/ple.py b/src/mcore_bridge/model/modules/ple.py
index 1121bd2..3de1e02 100644
--- a/src/mcore_bridge/model/modules/ple.py
+++ b/src/mcore_bridge/model/modules/ple.py
@@ -357,8 +357,6 @@ def forward(
         if thd:
             # SP keeps one global cu_seqlens copy per rank; normalize padded/offset
             # forms against the gathered total so cu indexes the full sequence.
-            # copy.copy (not dataclasses.replace) preserves dynamically attached
-            # fields such as `num_samples` that the data pipeline relies on.
             cu = self._normalize_cu_seqlens(getattr(packed_seq_params, 'cu_seqlens_q', None), hidden_states.shape[0])
             psp = copy.copy(packed_seq_params)
             psp.cu_seqlens_q = cu
@@ -382,13 +380,9 @@ def _forward_impl(
         """hidden_states: [s, b, nH] (bsh) or thd [T, 1, nH]; input_ids: [b, s] or [1, T]."""
         thd = packed_seq_params is not None and getattr(packed_seq_params, 'qkv_format', 'bshd') == 'thd'
         if thd:
-            num_samples = packed_seq_params.num_samples
-            # PackedSeqParams.max_seqlen_q is declared `int` in mcore and swift
-            # normalizes it to int, so `.item()` would raise AttributeError;
-            # tolerate a 0-d tensor from other callers.
-            max_seqlen_q = packed_seq_params.max_seqlen_q
-            max_len = int(max_seqlen_q.item() if torch.is_tensor(max_seqlen_q) else max_seqlen_q)
             cu = packed_seq_params.cu_seqlens_q
+            num_samples = cu.numel() - 1
+            max_len = int(packed_seq_params.max_seqlen_q)
             total = hidden_states.shape[0]
             hid = hidden_states.new_zeros((num_samples, max_len, hidden_states.shape[-1]))
             toks = input_ids.new_full((num_samples, max_len), self.ple_embedding.eos_token_id)