初始化项目，由ModelHub XC社区提供模型

Model: hxia7/Llama-3.2-1B-Block-FT
Source: Original Platform

.gitattributes

@@ -0,0 +1,35 @@
+*.7z filter=lfs diff=lfs merge=lfs -text
+*.arrow filter=lfs diff=lfs merge=lfs -text
+*.bin filter=lfs diff=lfs merge=lfs -text
+*.bz2 filter=lfs diff=lfs merge=lfs -text
+*.ckpt filter=lfs diff=lfs merge=lfs -text
+*.ftz filter=lfs diff=lfs merge=lfs -text
+*.gz filter=lfs diff=lfs merge=lfs -text
+*.h5 filter=lfs diff=lfs merge=lfs -text
+*.joblib filter=lfs diff=lfs merge=lfs -text
+*.lfs.* filter=lfs diff=lfs merge=lfs -text
+*.mlmodel filter=lfs diff=lfs merge=lfs -text
+*.model filter=lfs diff=lfs merge=lfs -text
+*.msgpack filter=lfs diff=lfs merge=lfs -text
+*.npy filter=lfs diff=lfs merge=lfs -text
+*.npz filter=lfs diff=lfs merge=lfs -text
+*.onnx filter=lfs diff=lfs merge=lfs -text
+*.ot filter=lfs diff=lfs merge=lfs -text
+*.parquet filter=lfs diff=lfs merge=lfs -text
+*.pb filter=lfs diff=lfs merge=lfs -text
+*.pickle filter=lfs diff=lfs merge=lfs -text
+*.pkl filter=lfs diff=lfs merge=lfs -text
+*.pt filter=lfs diff=lfs merge=lfs -text
+*.pth filter=lfs diff=lfs merge=lfs -text
+*.rar filter=lfs diff=lfs merge=lfs -text
+*.safetensors filter=lfs diff=lfs merge=lfs -text
+saved_model/**/* filter=lfs diff=lfs merge=lfs -text
+*.tar.* filter=lfs diff=lfs merge=lfs -text
+*.tar filter=lfs diff=lfs merge=lfs -text
+*.tflite filter=lfs diff=lfs merge=lfs -text
+*.tgz filter=lfs diff=lfs merge=lfs -text
+*.wasm filter=lfs diff=lfs merge=lfs -text
+*.xz filter=lfs diff=lfs merge=lfs -text
+*.zip filter=lfs diff=lfs merge=lfs -text
+*.zst filter=lfs diff=lfs merge=lfs -text
+*tfevents* filter=lfs diff=lfs merge=lfs -text

README.md

@@ -0,0 +1,160 @@
+---
+license: llama3.2
+base_model: meta-llama/Llama-3.2-1B
+tags:
+  - block-attention
+  - rag
+  - fine-tuned
+  - llama
+library_name: transformers
+pipeline_tag: text-generation
+---
+
+# Llama-3.2-1B-Block-FT
+
+Block-Attention fine-tuned [meta-llama/Llama-3.2-1B](https://huggingface.co/meta-llama/Llama-3.2-1B) for efficient RAG inference.
+
+## Overview
+
+This model is fine-tuned using the **Block-Attention** mechanism from [Block-Attention for Efficient Prefilling](https://openreview.net/forum?id=7zNYYbjpfW). Block-Attention divides the input context into independent blocks during the prefill phase, enabling **KV cache reuse** across different queries on the same documents — a key optimization for RAG serving.
+
+**Training Data Control Variable**: This model was fine-tuned on the **full 80K** samples from the Tulu3-Block-FT-RAG dataset. A companion Qwen3-8B model uses an 8K subset for comparison.
+
+## Evaluation Results
+
+### On Training Data (100 samples from Tulu3-Block-FT-RAG train set)
+
+*Note: These samples overlap with the training data, so absolute scores are inflated. The block vs full comparison remains valid.*
+
+| Model | EM Score | F1 Score |
+|-------|----------|----------|
+| meta-llama/Llama-3.2-1B (base) | 0.00% | 0.54% |
+| meta-llama/Llama-3.2-1B-Instruct | 2.00% | 50.62% |
+| **hxia7/Llama-3.2-1B-block-FT (full-attention)** | **14.00%** | **70.96%** |
+| **hxia7/Llama-3.2-1B-block-FT (block-attention)** | **17.00%** | **69.68%** |
+
+### On Unseen TriviaQA Validation Set (100 clean samples)
+
+*Questions and evidence passages from TriviaQA RC validation split, excluded from training data. Substr-EM checks whether the correct answer appears as a substring in the model's response.*
+
+| Model | Substr-EM | F1 Score |
+|-------|-----------|----------|
+| meta-llama/Llama-3.2-1B (base) | 56.00% | 12.51% |
+| meta-llama/Llama-3.2-1B-Instruct | 86.00% | 23.62% |
+| **hxia7/Llama-3.2-1B-block-FT (full-attention)** | **87.00%** | **26.59%** |
+| **hxia7/Llama-3.2-1B-block-FT (block-attention)** | **88.00%** | **27.53%** |
+
+Key observations:
+- Block-attention and full-attention produce **comparable results** across both evaluation sets, confirming the block-attention structure preserves quality.
+- On unseen data, block-FT outperforms the Instruct baseline in both Substr-EM (+2%) and F1 (+4%), demonstrating that RAG fine-tuning improves answer extraction quality even on out-of-distribution data.
+- The evidence passages from TriviaQA differ from the Contriever-retrieved passages used in training, making this a meaningful out-of-distribution test.
+
+## Block-Attention Mechanism
+
+In Block-Attention, the context is split into N blocks:
+- **Blocks 1..N-1 (document blocks)**: Use **local attention** — each block attends only to itself
+- **Block N (query block)**: Uses **global attention** — attends to all previous blocks
+
+This isolation allows document blocks' KV states to be computed once and reused across multiple queries.
+
+## Training Details
+
+- **Base Model**: meta-llama/Llama-3.2-1B
+- **Training Data**: Tulu3-Block-FT-RAG (80K samples, full dataset)
+- **Epochs**: 1
+- **Learning Rate**: 2e-5
+- **Optimizer**: AdamW (fused)
+- **Precision**: BF16
+- **DeepSpeed**: ZeRO Stage 2
+- **Loss Reduction**: sum (over non-masked tokens)
+
+During training, each sample produces two variants:
+1. **Full-attention** version (standard causal mask)
+2. **Block-attention** version (with `[Block-Attention]` prefix token and 4D block mask)
+
+Both variants contribute to the loss, teaching the model to handle both inference modes.
+
+## Inference
+
+### Block-Attention Inference (recommended for RAG)
+
+**Important**: Block-Attention uses a 4D attention mask `[1, 1, seq_len, seq_len]` during prefill. `model.generate()` only accepts 2D masks, so inference requires **manual prefill + autoregressive decode**:
+
+```python
+import torch
+from transformers import AutoTokenizer, AutoModelForCausalLM
+from src.data.block import build_attention_mask, convert_attention_mask_to_model_required
+
+model = AutoModelForCausalLM.from_pretrained("hxia7/Llama-3.2-1B-block-FT", torch_dtype=torch.bfloat16, device_map="auto")
+tokenizer = AutoTokenizer.from_pretrained("hxia7/Llama-3.2-1B-block-FT")
+
+blocks = [
+    "<|start_header_id|>system\nYou are an AI assistant. Below are reference documents.\n\n",
+    "- Title: Document 1\nContent of document 1...\n",
+    "- Title: Document 2\nContent of document 2...\n",
+    "Answer the question using the documents.\nQuestion: What is X?\n\n",
+]
+
+@torch.no_grad()
+def block_generate(model, tokenizer, blocks, max_new_tokens=128):
+    block_token_counts = []
+    all_ids = []
+    for b in blocks:
+        ids = tokenizer.encode(b, add_special_tokens=False)
+        all_ids.extend(ids)
+        block_token_counts.append(len(ids))
+
+    input_ids = torch.tensor([all_ids], dtype=torch.int64, device=model.device)
+    total_len = len(all_ids)
+
+    helper = torch.tril(torch.ones(total_len + 64, total_len + 64, dtype=torch.bool))
+    attn_mask = build_attention_mask(
+        local_attention_block_tokens=torch.tensor(block_token_counts[:-1], dtype=torch.long),
+        global_attention_block_tokens=torch.tensor(block_token_counts[-1], dtype=torch.long),
+        lower_triangular_matrix=helper,
+    )
+    attn_mask = convert_attention_mask_to_model_required(attn_mask)
+    attn_mask = attn_mask.unsqueeze(0).unsqueeze(0).to(model.device)
+
+    outputs = model(input_ids=input_ids, attention_mask=attn_mask, use_cache=True)
+    past_kv = outputs.past_key_values
+    next_token = torch.argmax(outputs.logits[:, -1, :], dim=-1, keepdim=True)
+
+    generated = []
+    for _ in range(max_new_tokens - 1):
+        if next_token.item() == tokenizer.eos_token_id:
+            break
+        generated.append(next_token.item())
+        outputs = model(input_ids=next_token, past_key_values=past_kv, use_cache=True)
+        past_kv = outputs.past_key_values
+        next_token = torch.argmax(outputs.logits[:, -1, :], dim=-1, keepdim=True)
+
+    if next_token.item() != tokenizer.eos_token_id:
+        generated.append(next_token.item())
+
+    return tokenizer.decode(generated, skip_special_tokens=True).strip()
+
+answer = block_generate(model, tokenizer, blocks)
+print(answer)
+```
+
+### Full-Attention Inference (standard)
+
+```python
+from transformers import AutoTokenizer, AutoModelForCausalLM
+
+model = AutoModelForCausalLM.from_pretrained("hxia7/Llama-3.2-1B-block-FT", torch_dtype=torch.bfloat16, device_map="auto")
+tokenizer = AutoTokenizer.from_pretrained("hxia7/Llama-3.2-1B-block-FT")
+
+prompt = "Your full RAG prompt here..."
+inputs = tokenizer(prompt, return_tensors="pt", truncation=True, max_length=3968).to(model.device)
+outputs = model.generate(**inputs, max_new_tokens=128, do_sample=False, pad_token_id=tokenizer.eos_token_id)
+answer = tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:], skip_special_tokens=True)
+print(answer)
+```
+
+## References
+
+- [Block-Attention for Efficient Prefilling](https://openreview.net/forum?id=7zNYYbjpfW)
+- [Official Implementation](https://github.com/TemporaryLoRA/Block-Attention)
+- Base model: [meta-llama/Llama-3.2-1B](https://huggingface.co/meta-llama/Llama-3.2-1B)

config.json

@@ -0,0 +1,36 @@
+{
+  "architectures": [
+    "LlamaForCausalLM"
+  ],
+  "attention_bias": false,
+  "attention_dropout": 0.0,
+  "bos_token_id": 128000,
+  "dtype": "bfloat16",
+  "eos_token_id": 128001,
+  "head_dim": 64,
+  "hidden_act": "silu",
+  "hidden_size": 2048,
+  "initializer_range": 0.02,
+  "intermediate_size": 8192,
+  "max_position_embeddings": 131072,
+  "mlp_bias": false,
+  "model_type": "llama",
+  "num_attention_heads": 32,
+  "num_hidden_layers": 16,
+  "num_key_value_heads": 8,
+  "pad_token_id": null,
+  "pretraining_tp": 1,
+  "rms_norm_eps": 1e-05,
+  "rope_parameters": {
+    "factor": 32.0,
+    "high_freq_factor": 4.0,
+    "low_freq_factor": 1.0,
+    "original_max_position_embeddings": 8192,
+    "rope_theta": 500000.0,
+    "rope_type": "llama3"
+  },
+  "tie_word_embeddings": true,
+  "transformers_version": "5.9.0",
+  "use_cache": false,
+  "vocab_size": 128256
+}

generation_config.json

@@ -0,0 +1,9 @@
+{
+  "_from_model_config": true,
+  "bos_token_id": 128000,
+  "do_sample": true,
+  "eos_token_id": 128001,
+  "temperature": 0.6,
+  "top_p": 0.9,
+  "transformers_version": "5.9.0"
+}

model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e9305f706586c40ae19a92aa660ff0f0a51d156f02d24c0af73bd4b354930c8e
+size 2996982344

special_tokens_map.json

@@ -0,0 +1,16 @@
+{
+  "bos_token": {
+    "content": "<|begin_of_text|>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "eos_token": {
+    "content": "<|end_of_text|>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  }
+}