Revert rope_type derivation from target

NOTE: this breaks compatibility with Meta's distributed DFlash GGUFs, as
the Q/K are stored in "NEOX" (rotated half) format, like in
transformers.

conversion/onyx.py

@@ -126,12 +126,6 @@ class OnyxAssistantModel(TextModel):
             self.gguf_writer.add_sliding_window_pattern([t == "sliding_attention" for t in h["layer_types"]])
 
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
-        # Invert transformers' permute_rope
-        if ".self_attn.q_proj." in name:
-            data_torch = _unpermute_for_rope(data_torch, int(self.hparams["num_attention_heads"]))
-        elif ".self_attn.k_proj." in name:
-            data_torch = _unpermute_for_rope(data_torch, int(self.hparams["num_key_value_heads"]))
-        elif ".self_attn.q_norm." in name or ".self_attn.k_norm." in name:
-            data_torch = _unpermute_for_rope(data_torch, 1)
-
+        # DFlash defaults to NEOX (rotate_half) rope, matching transformers HF layout for Q/K, QK-norms
+        # no permutation needed.
         yield (self.map_tensor_name(name), data_torch)

src/llama-graph.cpp

@@ -1350,10 +1350,7 @@ llm_graph_context::llm_graph_context(const llm_graph_params & params) :
     n_outputs        (params.n_outputs),
     n_ctx_orig       (cparams.n_ctx_orig_yarn),
     pooling_type     (cparams.pooling_type),
-    // DFlash: inherit rope type from the linked target
-    rope_type        ((arch == LLM_ARCH_DFLASH && cparams.ctx_other != nullptr)
-                      ? llama_get_model(cparams.ctx_other)->hparams.rope_type
-                      : hparams.rope_type),
+    rope_type        (hparams.rope_type),
     sched            (params.sched),
     backend_cpu      (params.backend_cpu),
     cvec             (params.cvec),