Update configuration_nandi.py

configuration_nandi.py

@@ -1,17 +1,3 @@
-# Copyright 2026 RTA AI Labs. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
 from transformers.configuration_utils import PretrainedConfig
 
 
@@ -97,26 +83,12 @@ class NandiConfig(PretrainedConfig):
         self.factorized_embedding = factorized_embedding
         self.embedding_rank = embedding_rank
         self.layer_sharing = layer_sharing
-        # Smoltron training loops over `layer_sharing_repeats` unconditionally
-        # (it does NOT check `layer_sharing`). Preserve the raw value here so
-        # the modeling code can honor it; the `layer_sharing` bool is now just
-        # metadata describing intent.
+
         self.layer_sharing_repeats = max(1, int(layer_sharing_repeats or 1))
         self.qk_norm = qk_norm
-        # `shared_kv` records that V was tied to K at pretraining time. In the
-        # HF model V is recomputed from `k_proj` at runtime (no `v_proj` module
-        # is materialised); see `NandiAttention.forward`.
+
         self.shared_kv = shared_kv
-        # `kv_cache_mode` controls the inference-time K/V cache strategy when
-        # `shared_kv=True`. Both modes produce identical outputs (numerical
-        # round-off only); they trade memory for compute:
-        #   "shared"  -> cache ONLY raw K (single tensor per layer). Each
-        #                decode step re-applies k_norm + RoPE to the full
-        #                cached raw K. Halves KV-cache memory.
-        #   "vanilla" -> cache post-norm post-RoPE K AND raw V (two tensors
-        #                per layer). k_norm + RoPE are applied only to the
-        #                current step's tokens. Standard HF behavior.
-        # Ignored when `shared_kv=False`. Defaults to "shared".
+
         if kv_cache_mode not in ("shared", "vanilla"):
             raise ValueError(
                 f"`kv_cache_mode` must be 'shared' or 'vanilla', got {kv_cache_mode!r}."