Merge remote-tracking branch 'origin/main' into cett3

Author: kaselby

configs/relullama_skip_causal_7b.json

@@ -0,0 +1,26 @@
+{
+    "_name_or_path": "SparseLLM/ReluLLaMA-7B",
+    "sparsity": 0.3,
+    "architectures": ["LlamaSkipConnectionForCausalLM"], 
+    "bos_token_id": 1, 
+    "eos_token_id": 2, 
+    "hidden_act": "relu",
+    "hidden_size": 4096, 
+    "initializer_range": 0.02, 
+    "intermediate_size": 11008, 
+    "max_length": 4096, 
+    "max_position_embeddings": 2048, 
+    "model_type": "llama-skip", 
+    "num_attention_heads": 32, 
+    "num_hidden_layers": 32, 
+    "num_key_value_heads": 32, 
+    "pad_token_id": 0, 
+    "pretraining_tp": 1, 
+    "rms_norm_eps": 1e-05, 
+    "rope_scaling": null, 
+    "tie_word_embeddings": false, 
+    "torch_dtype": "float32", 
+    "transformers_version": "4.31.0", 
+    "use_cache": true, 
+    "vocab_size": 32000
+}

sparse_transformers/csrc/sparse_mlp_op.cpp

@@ -148,7 +148,7 @@ torch::Tensor sparse_mlp_forward_cpu(
         auto gate_proj = combined_proj_block.narrow(1, 0, gate_size);  // [block_size, gate_size]
         auto up_proj = combined_proj_block.narrow(1, gate_size, gate_size);  // [block_size, gate_size]
 
-        gate_proj.sigmoid_();  // In-place sigmoid
+        gate_proj.relu_();  // In-place relu
         gate_proj.mul_(up_proj);  // In-place element-wise multiplication
 
         // Final projection to output dimension

src/models/llama/modelling_llama_skip.py

@@ -61,7 +61,7 @@ def _set_mlp_inference(self, config, layer_idx):
             config.intermediate_size,
             config.sparsity,
             config.mlp_bias,
-            "silu"
+            config.hidden_act
         )