init commit

README.md

@@ -1,3 +1,12 @@
 ---
 license: mit
+library_name: transformers
+base_model:
+- deepseek-ai/DeepSeek-R1-Distill-Qwen-14B
+base_model_relation: "adapter"
 ---
+
+This repo only contains the AttnGates' weights for deepseek-ai/DeepSeek-R1-Distill-Qwen-14B Model. It's only used for decoding. However, the current inference framework is unoptimized and only for accuracy tests. 
+
+[SeerAttention](https://arxiv.org/pdf/2410.13276) introduces learnable AttnGate modules to accelerate the computationally intensive prefill stage of long-context large language models (LLMs) via dynamic block-level sparsity. The AttnGates are trained in a parameter-efficient self-distillation framework, where they learn to mimic the block-wise attention patterns of the original frozen model, preserving its integrity while avoiding costly retraining. During inference, these gates generate block-sparse binary masks by applying threshold/TopK to their learned soft scores, enabling efficient computation through a custom block-sparse FlashAttention kernel. 
+

attn_gate_weights.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:abac6dde21f31e3e56e6888719f27eac93b34c3f871ff5bad8d70137d8e8b384
+size 100696226

config.json

@@ -0,0 +1,40 @@
+{
+  "_attn_implementation_autoset": true,
+  "_name_or_path": "/home/v-shumingguo/gsm_blob/distilled_models/DeepSeek-R1-Distill-Qwen-14BSFT-selfdata/bs16_steps500_selfdata",
+  "architectures": [
+    "SeerAttnQwen2ForCausalLM"
+  ],
+  "attention_dropout": 0.0,
+  "base_model": "deepseek-ai/DeepSeek-R1-Distill-Qwen-14B",
+  "bos_token_id": 151643,
+  "eos_token_id": 151643,
+  "hidden_act": "silu",
+  "hidden_size": 5120,
+  "initializer_range": 0.02,
+  "intermediate_size": 13824,
+  "max_position_embeddings": 131072,
+  "max_window_layers": 48,
+  "model_type": "qwen2",
+  "num_attention_heads": 40,
+  "num_hidden_layers": 48,
+  "num_key_value_heads": 8,
+  "rms_norm_eps": 1e-05,
+  "rope_scaling": null,
+  "rope_theta": 1000000.0,
+  "seerattn_gate_block_size": 64,
+  "seerattn_gate_hidden_size": 128,
+  "seerattn_gate_type": "Qavg_Kmaxminavg",
+  "seerattn_last_block_dense": true,
+  "seerattn_nz_ratio": 1.0,
+  "seerattn_sparsity_method": "threshold",
+  "seerattn_threshold": 0.0,
+  "sliding_window": 131072,
+  "tie_word_embeddings": false,
+  "torch_dtype": "bfloat16",
+  "transformers_version": "4.48.3",
+  "use_cache": true,
+  "use_decode_seerattn": true,
+  "use_prefill_seerattn": false,
+  "use_sliding_window": false,
+  "vocab_size": 151665
+}