Upload modeling_nsa.py with huggingface_hub

modeling_nsa.py

@@ -9,12 +9,7 @@ from transformers.generation.utils import GenerationMixin
 from transformers.modeling_outputs import CausalLMOutput
 
 from .configuration_nsa import NSAConfig
-_HAS_NSA = False
-try:
-    from .nsa.model.llama_block_nsa import LlamaBlockNSA as _VendorNSABlock
-    _HAS_NSA = True
-except Exception:
-    _VendorNSABlock = None  # type: ignore
+_HAS_NSA = False  # avoid nested vendor imports in HF dynamic loader
 
 
 class RMSNorm(nn.Module):
@@ -262,23 +257,7 @@ class NSATinyLM(nn.Module):
         import os as _os
         # Allow forcing simple fallback via env for integration tests
         _force_simple = _os.getenv('NSA_REMOTE_FORCE_SIMPLE', '0').lower() in ('1','true','yes')
-        if not _force_simple and _HAS_NSA and _VendorNSABlock is not None:
-            # Prefer vendored NSA block to match training semantics and map gate weights
-            self.blocks = nn.ModuleList([
-                _VendorNSABlock(
-                    dim=self.hidden_size,
-                    n_heads=self.num_attention_heads,
-                    n_kv_groups=self.n_kv_groups,
-                    d_k=self.d_k,
-                    d_v=self.d_v,
-                    l=self.l,
-                    d=self.d,
-                    l_sel=self.l_sel,
-                    n_sel=self.n_sel,
-                    w=self.w,
-                ) for _ in range(self.num_hidden_layers)
-            ])
-        elif not _force_simple:
+        if not _force_simple:
             # Fallback to embedded minimal NSA if vendor import failed
             self.blocks = nn.ModuleList([
                 NSABlockRemote(