Upload Llama4 Eagle3 dummy drafter for vLLM testing

README.md

@@ -1,55 +1,27 @@
----
-license: apache-2.0
-tags:
-- eagle3
-- speculative-decoding
-- llama4
-- vllm
-- testing
----
+# Llama4 Eagle Drafter Model (Test)
 
-# Llama4 Scout 17B Eagle3 Dummy Drafter
-
-This is a **dummy/test drafter model** for testing the Eagle3 speculative decoding implementation with Llama4 Scout 17B models in vLLM.
-
-⚠️ **WARNING**: This is not a real model and should not be used for inference. It contains random weights and is only for testing purposes.
+This is a test Eagle drafter model for Llama4 with proper configuration and vocabulary mappings.
 
 ## Model Details
-
-- **Architecture**: Llama4ForCausalLM (Eagle3 drafter variant)
-- **Target Model**: Llama4 Scout 17B  
-- **Base Model**: Llama4 architecture
-- **Hidden Size**: 2048
-- **Layers**: 1 (single decoder layer as per Eagle3 design)
-- **Vocabulary**: 128256 tokens
-- **Parameters**: ~322M
+- **Architecture**: Llama4ForCausalLM (Eagle draft variant)
+- **Hidden size**: 2048
+- **Layers**: 1 (single decoder layer for Eagle draft)
+- **Vocabulary**: 128256 tokens (Llama4)
+- **Includes**: d2t and t2d vocabulary mappings
 
 ## Configuration
-
-This model uses:
-- Eagle3 auxiliary hidden state combination
-- Llama4 layer structure with RoPE
-- SGLang-style weight naming (midlayer.*)
+- Uses standard Llama4 architecture
+- Includes Eagle auxiliary state configuration
+- Has vocabulary mapping tensors (d2t/t2d) for draft-to-target conversion
+- Extended context support (262k max position embeddings)
 
 ## Usage
+This model is for testing Eagle speculative decoding with Llama4 in vLLM:
 
-This model is designed specifically for testing the vLLM Eagle3 implementation:
-
-```python
-# Use with vLLM for testing Eagle3 speculative decoding with Llama4 Scout
-vllm serve RedHatAI/Llama-4-Scout-17B-16E-Instruct-quantized.w4a16 \
-    --speculative-config '{"method": "eagle3", "model": "nm-testing/llama4-scout-17b-eagle3-dummy-drafter", ...}'
+```bash
+vllm serve <llama4-target-model> \
+    --speculative-config '{"method": "eagle", "model": "nm-testing/llama4-eagle-drafter", ...}'
 ```
 
-## Testing Purpose Only
-
-This model:
-- Contains random weights
-- Is not trained on any data
-- Should not be used for actual inference
-- Is only for vLLM development and testing
-
-## Related
-
-- vLLM: https://github.com/vllm-project/vllm
-- Eagle3: Speculative decoding method
+## Testing Purpose
+This model contains random weights and is only for vLLM Eagle implementation testing.

config.json

@@ -5,37 +5,17 @@
   "model_type": "llama4_text",
   "hidden_size": 2048,
   "intermediate_size": 5632,
-  "intermediate_size_mlp": 5632,
   "num_hidden_layers": 1,
   "num_attention_heads": 16,
   "num_key_value_heads": 8,
   "head_dim": 128,
   "vocab_size": 128256,
-  "draft_vocab_size": 128256,
   "attention_bias": false,
   "attention_dropout": 0.0,
-  "attn_temperature_tuning": false,
-  "cross_layer_interval": 1,
-  "ffn_dim_multiplier": null,
   "mlp_bias": false,
-  "multi_head_axial": false,
-  "original_multi_head": false,
-  "scale_attention_scores": false,
   "tie_word_embeddings": false,
   "use_cache": true,
-  "use_memory_efficient_attention": false,
   "use_qk_norm": false,
-  "no_rope_layers": [
-    1
-  ],
-  "attention_chunk_size": 8192,
-  "interleave_moe_layer_step": 0,
-  "moe_layers": [],
-  "yoco_global_kv_layer": null,
-  "yoco_local_kv_layer": null,
-  "num_local_experts": null,
-  "num_experts_per_tok": null,
-  "router_aux_loss_coef": null,
   "rope_theta": 500000.0,
   "rope_scaling": {
     "factor": 8.0,
@@ -50,7 +30,6 @@
     "mscale_all_dim": 1.0
   },
   "max_position_embeddings": 262144,
-  "logit_scale": 1.0,
   "rms_norm_eps": 1e-05,
   "hidden_act": "silu",
   "initializer_range": 0.02,