Fix correctness bugs across multiple model files

1. cohere.py:347-348 - Fixed wrong variable names in QK normalization.
   Used `Q`/`K` but variables were named `Qn`/`Kn`. This caused NameError
   when `use_qk_norm=True` (e.g., c4ai-command-r-plus models).

2. cohere.py:482 - Fixed wrong object reference in inference loop.
   Used `self.mlp` but should be `decoder_layer.mlp` since we're
   iterating through decoder layers. Caused AttributeError during inference.

3. falcon_h1.py:459,461 - Fixed wrong attribute names in inference path.
   Used `post_attention_layernorm` and `mlp` but Falcon H1 uses
   `pre_ff_layernorm` and `feed_forward`. Caused AttributeError during generation.

4. qwen3_moe.py:210 - Fixed wrong module path with incorrect capitalization.
   Used `transformers.models.Qwen3Moe` but should be `transformers.models.qwen3_moe`.
   Caused AttributeError when patching rotary embeddings.

5. qwen3_moe.py:239 - Fixed wrong model_patcher class.
   Used `FastQwen3Model` but should be `FastQwen3MoeModel` for MoE models.
   Caused incorrect patching for Qwen3 MoE models.

6. hf_hub.py:21-22 - Fixed floor division and missing return for billion values.
   Used `//` instead of `/` for millions, and had no return for values >= 1B.
   Caused incorrect formatting and None return for large numbers.

7. save.py:550 - Fixed self-assignment that did nothing.
   `sharded_ram_usage = sharded_ram_usage` should be `= max_shard_size`.
   Caused integer shard sizes to be ignored.

8. rl.py:562-567 - Fixed orphan string not included in length_check.
   The elif branch for max_seq_length validation was a standalone string
   expression, not concatenated to length_check. Caused silent skip of
   the max_seq_length > model_max_seq_length warning.

9. granite.py:49-52 - Fixed wrong model name and version in error message.
   Said "Gemma2" and "4.42.3" but should be "Granite" and "4.45.0".

unsloth/models/cohere.py

@@ -344,8 +344,8 @@ def CohereAttention_fast_forward_inference(
     Kn = Kn.view(bsz, 1, n_kv_heads, head_dim).transpose(1, 2)
     Vn = Vn.view(bsz, 1, n_kv_heads, head_dim).transpose(1, 2)
     if self.use_qk_norm:
-        Q = fast_layernorm_inference(self.q_norm, Q, self.q_norm_out_weight)
-        K = fast_layernorm_inference(self.k_norm, K, self.k_norm_out_weight)
+        Qn = fast_layernorm_inference(self.q_norm, Qn, self.q_norm_out_weight)
+        Kn = fast_layernorm_inference(self.k_norm, Kn, self.k_norm_out_weight)
 
     # cos, sin = self.rotary_emb(Vn, seq_len = kv_seq_len)
     # Qn, Kn = inplace_rope_embedding(Qn, Kn, cos, sin, position_ids)
@@ -479,7 +479,7 @@ def CohereModel_fast_forward_inference(
             )
         )
 
-        hidden_states_mlp = fast_swiglu_inference(self.mlp, hidden_states)
+        hidden_states_mlp = fast_swiglu_inference(decoder_layer.mlp, hidden_states)
         residual += hidden_states_attention
         residual += hidden_states_mlp
         hidden_states = residual

unsloth/models/falcon_h1.py

@@ -456,9 +456,9 @@ def FalconH1DecoderLayer_fast_forward(
         # Fully Connected
         residual = hidden_states
         hidden_states = fast_rms_layernorm_inference(
-            self.post_attention_layernorm, hidden_states
+            self.pre_ff_layernorm, hidden_states
         )
-        hidden_states = fast_swiglu_inference(self.mlp, hidden_states)
+        hidden_states = fast_swiglu_inference(self.feed_forward, hidden_states)
         hidden_states += residual
     else:
         residual = hidden_states

unsloth/models/granite.py

@@ -46,9 +46,9 @@ except:
     transformers_version = Version(transformers_version)
     if not transformers_version >= Version("4.45.0"):
         raise ImportError(
-            f"Unsloth: Your transformers version of {transformers_version} does not support Gemma2.\n"
-            f"The minimum required version is 4.42.3.\n"
-            f'Try `pip install --upgrade "transformers>=4.42.3"`\n'
+            f"Unsloth: Your transformers version of {transformers_version} does not support Granite.\n"
+            f"The minimum required version is 4.45.0.\n"
+            f'Try `pip install --upgrade "transformers>=4.45.0"`\n'
             f"to obtain the latest transformers build, then restart this session."
         )
 

unsloth/models/qwen3_moe.py

@@ -207,7 +207,7 @@ class FastQwen3MoeModel(FastQwen3Model):
         # https://github.com/huggingface/transformers/blob/v4.37.2/src/transformers/models/llama/modeling_llama.py\
         import transformers.models.qwen3_moe.modeling_qwen3_moe
 
-        transformers.models.Qwen3Moe.modeling_qwen3_moe.Qwen3MoeRotaryEmbedding = (
+        transformers.models.qwen3_moe.modeling_qwen3_moe.Qwen3MoeRotaryEmbedding = (
             LlamaRotaryEmbedding
         )
         return
@@ -236,7 +236,7 @@ class FastQwen3MoeModel(FastQwen3Model):
             device_map = device_map,
             rope_scaling = rope_scaling,
             fix_tokenizer = fix_tokenizer,
-            model_patcher = FastQwen3Model,
+            model_patcher = FastQwen3MoeModel,
             tokenizer_name = tokenizer_name,
             trust_remote_code = trust_remote_code,
             **kwargs,

unsloth/models/rl.py

@@ -559,8 +559,12 @@ def _patch_trl_rl_trainers(trainer_file = "grpo_trainer"):
             "    if args_max_seq_length is None and model_max_seq_length is not None:\n"
             "        max_seq_length = model.max_seq_length\n"
             "        if hasattr(args, 'max_seq_length'): args.max_seq_length = max_seq_length\n"
+            "    elif args_max_seq_length is not None and model_max_seq_length is not None:\n"
+            "        if args_max_seq_length > model_max_seq_length:\n"
+            "            print('Unsloth: You set `max_seq_length` as ' + str(args_max_seq_length) + ' but '\n"
+            "                   'the maximum the model supports is ' + str(model_max_seq_length) + '. We shall reduce it.')\n"
+            "            args.max_seq_length = model_max_seq_length\n"
         )
-        "    elif args_max_seq_length is not None and model_max_seq_length is not None:\n" "        if args_max_seq_length > model_max_seq_length:\n" "            print('Unsloth: You set `max_seq_length` as ' + str(args_max_seq_length) + ' but \n" "                   the maximum the model supports is ' + str(model_max_seq_length) + '. We shall reduce it.')\n" "            args.max_seq_length = model_max_seq_length\n"
         extra_args += length_check
 
         # At this point max_seq_length might be set, but trl is moving to max_length

unsloth/save.py

@@ -547,7 +547,7 @@ def unsloth_save_model(
         elif mb_found:
             sharded_ram_usage = int(mb_found.group(1)) * 1024 * 1024
     elif type(max_shard_size) is int:
-        sharded_ram_usage = sharded_ram_usage
+        sharded_ram_usage = max_shard_size
 
     # Switch to our fast saving modules if it's a slow PC!
     n_cpus = psutil.cpu_count(logical = False)

unsloth/utils/hf_hub.py

@@ -19,7 +19,9 @@ def formatted_int(value: int) -> str:
     elif value < MILLION:
         return f"{float(value) / 1000:,.1f}K"
     elif value < BILLION:
-        return f"{float(value) // 1000000:,.1f}M"
+        return f"{float(value) / 1000000:,.1f}M"
+    else:
+        return f"{float(value) / 1000000000:,.1f}B"
 
 
 def get_model_info(