Add LLaMA-2-13B SVD-LLM / SVD-LLM V2 baselines (healed 60M and unhealed)

README.md

@@ -80,3 +80,16 @@ License. Do not assume one subdirectory's terms apply to another.
 requires rebuilding the architecture with the layer classes shipped in each directory, using
 that directory's `ranks.json`. `AutoModelForCausalLM.from_pretrained()` will **not** work.
 Each subdirectory's README carries a complete loading recipe.
+
+## LLaMA-2-13B SVD baselines (healed vs unhealed)
+
+13B counterparts of the 7B SVD baselines at the matched 13B budget (7.7403B, 59.47% kept —
+same as TR-3 13B), KD-healed with the identical 60M-token protocol. `*-unhealed` dirs hold the
+raw compression weights (no distillation) so the healing delta is directly measurable:
+
+| Directory | Stage | Method | Params |
+|---|---|---|---:|
+| [`svd-llm-13b/`](./svd-llm-13b) | KD-healed | SVD-LLM | 7.738B |
+| [`svd-llm-13b-unhealed/`](./svd-llm-13b-unhealed) | raw compression | SVD-LLM | 7.738B |
+| [`svd-llm-v2-13b/`](./svd-llm-v2-13b) | KD-healed | SVD-LLM V2 ‡ | 7.7403B |
+| [`svd-llm-v2-13b-unhealed/`](./svd-llm-v2-13b-unhealed) | raw compression | SVD-LLM V2 ‡ | 7.7403B |

svd-llm-13b-unhealed/README.md

@@ -0,0 +1,16 @@
+---
+license: llama2
+base_model: NousResearch/Llama-2-13b-chat-hf
+tags: [svd, low-rank, model-compression, llama2]
+---
+
+# SVD-LLM @ 7.74B (59.47% kept of LLaMA-2-13B-chat), UNHEALED
+
+SVD-LLM (ICLR 2025): truncation-aware data whitening + closed-form parameter update, official code (@ 7538cca), uniform 41.58% removal across all 7 projection types of all 40 layers.
+
+- **UNHEALED checkpoint**: raw post-training compression weights, NO distillation. The KD-healed
+  counterpart is in `svd-llm-13b/`. This dir exists so the healing delta is
+  measurable; use it as the compression-stage artifact, not as a deployable model.
+- Same `ranks.json` as the healed dir; weights differ (see `provenance.json` SHA256).
+
+Load exactly as in the sibling dir (`lowrank_linear.py`).

svd-llm-13b-unhealed/config.json

@@ -0,0 +1,26 @@
+{
+  "_name_or_path": null,
+  "architectures": [
+    "LlamaForCausalLM"
+  ],
+  "bos_token_id": 1,
+  "eos_token_id": 2,
+  "hidden_act": "silu",
+  "hidden_size": 5120,
+  "initializer_range": 0.02,
+  "intermediate_size": 13824,
+  "max_position_embeddings": 4096,
+  "model_type": "llama",
+  "num_attention_heads": 40,
+  "num_hidden_layers": 40,
+  "num_key_value_heads": 40,
+  "pad_token_id": 0,
+  "pretraining_tp": 1,
+  "rms_norm_eps": 1e-05,
+  "rope_scaling": null,
+  "tie_word_embeddings": false,
+  "torch_dtype": "float16",
+  "transformers_version": "4.31.0.dev0",
+  "use_cache": true,
+  "vocab_size": 32000
+}

svd-llm-13b-unhealed/lowrank_linear.py

@@ -0,0 +1,83 @@
+"""Minimal standalone loader for the SVD-baseline compressed LLaMA-2-7B checkpoints.
+
+No dependency beyond torch + transformers. The state_dict stores every compressed
+projection as two factors of a LowRankLinear (A: out x k, B: k x in); ranks.json
+holds the per-layer integer rank of each projection. Untouched parameters
+(embeddings, norms) keep their stock transformers names.
+
+Usage:
+    import json, torch
+    from lowrank_linear import build
+    model = build(
+        base_id="NousResearch/Llama-2-7b-chat-hf",
+        ranks=json.load(open("ranks.json")),
+        state_dict_path="state_dict_fp16.pt",
+        device="cuda", dtype=torch.float16,
+    )
+    tokenizer = AutoTokenizer.from_pretrained("NousResearch/Llama-2-7b-chat-hf")
+    out = model.generate(tokenizer("2+2=", return_tensors="pt").to("cuda").input_ids,
+                         max_new_tokens=16, do_sample=False)
+    print(tokenizer.decode(out[0]))
+"""
+
+from __future__ import annotations
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+
+class LowRankLinear(nn.Module):
+    """y = A (B x).  A: out x r, B: r x in.  (LLaMA projections carry no bias.)"""
+
+    def __init__(self, A: torch.Tensor, B: torch.Tensor):
+        super().__init__()
+        self.A = nn.Parameter(A)
+        self.B = nn.Parameter(B)
+
+    def forward(self, x):
+        return F.linear(F.linear(x, self.B), self.A)
+
+    @torch.no_grad()
+    def materialize_(self) -> nn.Linear:
+        """Replace this module with the dense W = A @ B (inference only; never call
+        while training -- it cuts the gradient path)."""
+        W = (self.A @ self.B).to(self.A.dtype)
+        dense = nn.Linear(self.B.shape[1], self.A.shape[0], bias=False,
+                          device=W.device, dtype=W.dtype)
+        dense.weight.data = W
+        return dense
+
+
+def build(base_id: str, ranks: dict, state_dict_path: str,
+          device: str = "cuda", dtype=torch.float16):
+    from transformers import AutoModelForCausalLM
+
+    model = AutoModelForCausalLM.from_pretrained(base_id, torch_dtype=dtype)
+    search = {}
+    for name, mod in model.named_modules():
+        if isinstance(mod, nn.Linear):
+            search[name] = mod
+    for path, k in ranks.items():
+        out_f, in_f = search[path].weight.shape
+        model_mod = model
+        for p in path.split(".")[:-1]:
+            model_mod = getattr(model_mod, p)
+        setattr(model_mod, path.split(".")[-1],
+                LowRankLinear(torch.zeros(out_f, int(k), dtype=dtype),
+                              torch.zeros(int(k), in_f, dtype=dtype)))
+    sd = torch.load(state_dict_path, map_location="cpu", weights_only=True)
+    model.load_state_dict(sd, strict=True)
+    del sd
+    return model.to(device)
+
+
+def materialize_all(model):
+    """Fuse every LowRankLinear back into a dense nn.Linear in place."""
+    for name, mod in list(model.named_modules()):
+        if isinstance(mod, LowRankLinear):
+            parent = model
+            for p in name.split(".")[:-1]:
+                parent = getattr(parent, p)
+            setattr(parent, name.split(".")[-1], mod.materialize_())
+    return model

svd-llm-13b-unhealed/provenance.json

@@ -0,0 +1,7 @@
+{
+ "kind": "unhealed SVD compression (TR3-format conversion, verified pure re-keying)",
+ "source_ckpt": "/data1/rkp00022/zixuan/projects/TR3_main/checkpoints/svdllm1_llama2_13b_chat_r04158/NousResearch_Llama_2_13b_chat_hf_whitening_then_update_0.5842.pt",
+ "converted_state_dict": "/data1/rkp00022/zixuan/projects/TR3_main/scratch/hf_push_13b/svd-llm-13b-unhealed/state_dict_fp16.pt",
+ "state_dict_sha256": "5625041bed874f21664b6a2bff3ce4722612302630028c9b60ac7f2cb8d7c605",
+ "healed_counterpart": "svd-llm-13b/state_dict_fp16.pt"
+}

svd-llm-13b-unhealed/ranks.json

@@ -0,0 +1,282 @@
+{
+ "model.layers.0.self_attn.q_proj": 1495,
+ "model.layers.0.self_attn.k_proj": 1495,
+ "model.layers.0.self_attn.v_proj": 1495,
+ "model.layers.0.self_attn.o_proj": 1495,
+ "model.layers.0.mlp.gate_proj": 2182,
+ "model.layers.0.mlp.up_proj": 2182,
+ "model.layers.0.mlp.down_proj": 2182,
+ "model.layers.1.self_attn.q_proj": 1495,
+ "model.layers.1.self_attn.k_proj": 1495,
+ "model.layers.1.self_attn.v_proj": 1495,
+ "model.layers.1.self_attn.o_proj": 1495,
+ "model.layers.1.mlp.gate_proj": 2182,
+ "model.layers.1.mlp.up_proj": 2182,
+ "model.layers.1.mlp.down_proj": 2182,
+ "model.layers.2.self_attn.q_proj": 1495,
+ "model.layers.2.self_attn.k_proj": 1495,
+ "model.layers.2.self_attn.v_proj": 1495,
+ "model.layers.2.self_attn.o_proj": 1495,
+ "model.layers.2.mlp.gate_proj": 2182,
+ "model.layers.2.mlp.up_proj": 2182,
+ "model.layers.2.mlp.down_proj": 2182,
+ "model.layers.3.self_attn.q_proj": 1495,
+ "model.layers.3.self_attn.k_proj": 1495,
+ "model.layers.3.self_attn.v_proj": 1495,
+ "model.layers.3.self_attn.o_proj": 1495,
+ "model.layers.3.mlp.gate_proj": 2182,
+ "model.layers.3.mlp.up_proj": 2182,
+ "model.layers.3.mlp.down_proj": 2182,
+ "model.layers.4.self_attn.q_proj": 1495,
+ "model.layers.4.self_attn.k_proj": 1495,
+ "model.layers.4.self_attn.v_proj": 1495,
+ "model.layers.4.self_attn.o_proj": 1495,
+ "model.layers.4.mlp.gate_proj": 2182,
+ "model.layers.4.mlp.up_proj": 2182,
+ "model.layers.4.mlp.down_proj": 2182,
+ "model.layers.5.self_attn.q_proj": 1495,
+ "model.layers.5.self_attn.k_proj": 1495,
+ "model.layers.5.self_attn.v_proj": 1495,
+ "model.layers.5.self_attn.o_proj": 1495,
+ "model.layers.5.mlp.gate_proj": 2182,
+ "model.layers.5.mlp.up_proj": 2182,
+ "model.layers.5.mlp.down_proj": 2182,
+ "model.layers.6.self_attn.q_proj": 1495,
+ "model.layers.6.self_attn.k_proj": 1495,
+ "model.layers.6.self_attn.v_proj": 1495,
+ "model.layers.6.self_attn.o_proj": 1495,
+ "model.layers.6.mlp.gate_proj": 2182,
+ "model.layers.6.mlp.up_proj": 2182,
+ "model.layers.6.mlp.down_proj": 2182,
+ "model.layers.7.self_attn.q_proj": 1495,
+ "model.layers.7.self_attn.k_proj": 1495,
+ "model.layers.7.self_attn.v_proj": 1495,
+ "model.layers.7.self_attn.o_proj": 1495,
+ "model.layers.7.mlp.gate_proj": 2182,
+ "model.layers.7.mlp.up_proj": 2182,
+ "model.layers.7.mlp.down_proj": 2182,
+ "model.layers.8.self_attn.q_proj": 1495,
+ "model.layers.8.self_attn.k_proj": 1495,
+ "model.layers.8.self_attn.v_proj": 1495,
+ "model.layers.8.self_attn.o_proj": 1495,
+ "model.layers.8.mlp.gate_proj": 2182,
+ "model.layers.8.mlp.up_proj": 2182,
+ "model.layers.8.mlp.down_proj": 2182,
+ "model.layers.9.self_attn.q_proj": 1495,
+ "model.layers.9.self_attn.k_proj": 1495,
+ "model.layers.9.self_attn.v_proj": 1495,
+ "model.layers.9.self_attn.o_proj": 1495,
+ "model.layers.9.mlp.gate_proj": 2182,
+ "model.layers.9.mlp.up_proj": 2182,
+ "model.layers.9.mlp.down_proj": 2182,
+ "model.layers.10.self_attn.q_proj": 1495,
+ "model.layers.10.self_attn.k_proj": 1495,
+ "model.layers.10.self_attn.v_proj": 1495,
+ "model.layers.10.self_attn.o_proj": 1495,
+ "model.layers.10.mlp.gate_proj": 2182,
+ "model.layers.10.mlp.up_proj": 2182,
+ "model.layers.10.mlp.down_proj": 2182,
+ "model.layers.11.self_attn.q_proj": 1495,
+ "model.layers.11.self_attn.k_proj": 1495,
+ "model.layers.11.self_attn.v_proj": 1495,
+ "model.layers.11.self_attn.o_proj": 1495,
+ "model.layers.11.mlp.gate_proj": 2182,
+ "model.layers.11.mlp.up_proj": 2182,
+ "model.layers.11.mlp.down_proj": 2182,
+ "model.layers.12.self_attn.q_proj": 1495,
+ "model.layers.12.self_attn.k_proj": 1495,
+ "model.layers.12.self_attn.v_proj": 1495,
+ "model.layers.12.self_attn.o_proj": 1495,
+ "model.layers.12.mlp.gate_proj": 2182,
+ "model.layers.12.mlp.up_proj": 2182,
+ "model.layers.12.mlp.down_proj": 2182,
+ "model.layers.13.self_attn.q_proj": 1495,
+ "model.layers.13.self_attn.k_proj": 1495,
+ "model.layers.13.self_attn.v_proj": 1495,
+ "model.layers.13.self_attn.o_proj": 1495,
+ "model.layers.13.mlp.gate_proj": 2182,
+ "model.layers.13.mlp.up_proj": 2182,
+ "model.layers.13.mlp.down_proj": 2182,
+ "model.layers.14.self_attn.q_proj": 1495,
+ "model.layers.14.self_attn.k_proj": 1495,
+ "model.layers.14.self_attn.v_proj": 1495,
+ "model.layers.14.self_attn.o_proj": 1495,
+ "model.layers.14.mlp.gate_proj": 2182,
+ "model.layers.14.mlp.up_proj": 2182,
+ "model.layers.14.mlp.down_proj": 2182,
+ "model.layers.15.self_attn.q_proj": 1495,
+ "model.layers.15.self_attn.k_proj": 1495,
+ "model.layers.15.self_attn.v_proj": 1495,
+ "model.layers.15.self_attn.o_proj": 1495,
+ "model.layers.15.mlp.gate_proj": 2182,
+ "model.layers.15.mlp.up_proj": 2182,
+ "model.layers.15.mlp.down_proj": 2182,
+ "model.layers.16.self_attn.q_proj": 1495,
+ "model.layers.16.self_attn.k_proj": 1495,
+ "model.layers.16.self_attn.v_proj": 1495,
+ "model.layers.16.self_attn.o_proj": 1495,
+ "model.layers.16.mlp.gate_proj": 2182,
+ "model.layers.16.mlp.up_proj": 2182,
+ "model.layers.16.mlp.down_proj": 2182,
+ "model.layers.17.self_attn.q_proj": 1495,
+ "model.layers.17.self_attn.k_proj": 1495,
+ "model.layers.17.self_attn.v_proj": 1495,
+ "model.layers.17.self_attn.o_proj": 1495,
+ "model.layers.17.mlp.gate_proj": 2182,
+ "model.layers.17.mlp.up_proj": 2182,
+ "model.layers.17.mlp.down_proj": 2182,
+ "model.layers.18.self_attn.q_proj": 1495,
+ "model.layers.18.self_attn.k_proj": 1495,
+ "model.layers.18.self_attn.v_proj": 1495,
+ "model.layers.18.self_attn.o_proj": 1495,
+ "model.layers.18.mlp.gate_proj": 2182,
+ "model.layers.18.mlp.up_proj": 2182,
+ "model.layers.18.mlp.down_proj": 2182,
+ "model.layers.19.self_attn.q_proj": 1495,
+ "model.layers.19.self_attn.k_proj": 1495,
+ "model.layers.19.self_attn.v_proj": 1495,
+ "model.layers.19.self_attn.o_proj": 1495,
+ "model.layers.19.mlp.gate_proj": 2182,
+ "model.layers.19.mlp.up_proj": 2182,
+ "model.layers.19.mlp.down_proj": 2182,
+ "model.layers.20.self_attn.q_proj": 1495,
+ "model.layers.20.self_attn.k_proj": 1495,
+ "model.layers.20.self_attn.v_proj": 1495,
+ "model.layers.20.self_attn.o_proj": 1495,
+ "model.layers.20.mlp.gate_proj": 2182,
+ "model.layers.20.mlp.up_proj": 2182,
+ "model.layers.20.mlp.down_proj": 2182,
+ "model.layers.21.self_attn.q_proj": 1495,
+ "model.layers.21.self_attn.k_proj": 1495,
+ "model.layers.21.self_attn.v_proj": 1495,
+ "model.layers.21.self_attn.o_proj": 1495,
+ "model.layers.21.mlp.gate_proj": 2182,
+ "model.layers.21.mlp.up_proj": 2182,
+ "model.layers.21.mlp.down_proj": 2182,
+ "model.layers.22.self_attn.q_proj": 1495,
+ "model.layers.22.self_attn.k_proj": 1495,
+ "model.layers.22.self_attn.v_proj": 1495,
+ "model.layers.22.self_attn.o_proj": 1495,
+ "model.layers.22.mlp.gate_proj": 2182,
+ "model.layers.22.mlp.up_proj": 2182,
+ "model.layers.22.mlp.down_proj": 2182,
+ "model.layers.23.self_attn.q_proj": 1495,
+ "model.layers.23.self_attn.k_proj": 1495,
+ "model.layers.23.self_attn.v_proj": 1495,
+ "model.layers.23.self_attn.o_proj": 1495,
+ "model.layers.23.mlp.gate_proj": 2182,
+ "model.layers.23.mlp.up_proj": 2182,
+ "model.layers.23.mlp.down_proj": 2182,
+ "model.layers.24.self_attn.q_proj": 1495,
+ "model.layers.24.self_attn.k_proj": 1495,
+ "model.layers.24.self_attn.v_proj": 1495,
+ "model.layers.24.self_attn.o_proj": 1495,
+ "model.layers.24.mlp.gate_proj": 2182,
+ "model.layers.24.mlp.up_proj": 2182,
+ "model.layers.24.mlp.down_proj": 2182,
+ "model.layers.25.self_attn.q_proj": 1495,
+ "model.layers.25.self_attn.k_proj": 1495,
+ "model.layers.25.self_attn.v_proj": 1495,
+ "model.layers.25.self_attn.o_proj": 1495,
+ "model.layers.25.mlp.gate_proj": 2182,
+ "model.layers.25.mlp.up_proj": 2182,
+ "model.layers.25.mlp.down_proj": 2182,
+ "model.layers.26.self_attn.q_proj": 1495,
+ "model.layers.26.self_attn.k_proj": 1495,
+ "model.layers.26.self_attn.v_proj": 1495,
+ "model.layers.26.self_attn.o_proj": 1495,
+ "model.layers.26.mlp.gate_proj": 2182,
+ "model.layers.26.mlp.up_proj": 2182,
+ "model.layers.26.mlp.down_proj": 2182,
+ "model.layers.27.self_attn.q_proj": 1495,
+ "model.layers.27.self_attn.k_proj": 1495,
+ "model.layers.27.self_attn.v_proj": 1495,
+ "model.layers.27.self_attn.o_proj": 1495,
+ "model.layers.27.mlp.gate_proj": 2182,
+ "model.layers.27.mlp.up_proj": 2182,
+ "model.layers.27.mlp.down_proj": 2182,
+ "model.layers.28.self_attn.q_proj": 1495,
+ "model.layers.28.self_attn.k_proj": 1495,
+ "model.layers.28.self_attn.v_proj": 1495,
+ "model.layers.28.self_attn.o_proj": 1495,
+ "model.layers.28.mlp.gate_proj": 2182,
+ "model.layers.28.mlp.up_proj": 2182,
+ "model.layers.28.mlp.down_proj": 2182,
+ "model.layers.29.self_attn.q_proj": 1495,
+ "model.layers.29.self_attn.k_proj": 1495,
+ "model.layers.29.self_attn.v_proj": 1495,
+ "model.layers.29.self_attn.o_proj": 1495,
+ "model.layers.29.mlp.gate_proj": 2182,
+ "model.layers.29.mlp.up_proj": 2182,
+ "model.layers.29.mlp.down_proj": 2182,
+ "model.layers.30.self_attn.q_proj": 1495,
+ "model.layers.30.self_attn.k_proj": 1495,
+ "model.layers.30.self_attn.v_proj": 1495,
+ "model.layers.30.self_attn.o_proj": 1495,
+ "model.layers.30.mlp.gate_proj": 2182,
+ "model.layers.30.mlp.up_proj": 2182,
+ "model.layers.30.mlp.down_proj": 2182,
+ "model.layers.31.self_attn.q_proj": 1495,
+ "model.layers.31.self_attn.k_proj": 1495,
+ "model.layers.31.self_attn.v_proj": 1495,
+ "model.layers.31.self_attn.o_proj": 1495,
+ "model.layers.31.mlp.gate_proj": 2182,
+ "model.layers.31.mlp.up_proj": 2182,
+ "model.layers.31.mlp.down_proj": 2182,
+ "model.layers.32.self_attn.q_proj": 1495,
+ "model.layers.32.self_attn.k_proj": 1495,
+ "model.layers.32.self_attn.v_proj": 1495,
+ "model.layers.32.self_attn.o_proj": 1495,
+ "model.layers.32.mlp.gate_proj": 2182,
+ "model.layers.32.mlp.up_proj": 2182,
+ "model.layers.32.mlp.down_proj": 2182,
+ "model.layers.33.self_attn.q_proj": 1495,
+ "model.layers.33.self_attn.k_proj": 1495,
+ "model.layers.33.self_attn.v_proj": 1495,
+ "model.layers.33.self_attn.o_proj": 1495,
+ "model.layers.33.mlp.gate_proj": 2182,
+ "model.layers.33.mlp.up_proj": 2182,
+ "model.layers.33.mlp.down_proj": 2182,
+ "model.layers.34.self_attn.q_proj": 1495,
+ "model.layers.34.self_attn.k_proj": 1495,
+ "model.layers.34.self_attn.v_proj": 1495,
+ "model.layers.34.self_attn.o_proj": 1495,
+ "model.layers.34.mlp.gate_proj": 2182,
+ "model.layers.34.mlp.up_proj": 2182,
+ "model.layers.34.mlp.down_proj": 2182,
+ "model.layers.35.self_attn.q_proj": 1495,
+ "model.layers.35.self_attn.k_proj": 1495,
+ "model.layers.35.self_attn.v_proj": 1495,
+ "model.layers.35.self_attn.o_proj": 1495,
+ "model.layers.35.mlp.gate_proj": 2182,
+ "model.layers.35.mlp.up_proj": 2182,
+ "model.layers.35.mlp.down_proj": 2182,
+ "model.layers.36.self_attn.q_proj": 1495,
+ "model.layers.36.self_attn.k_proj": 1495,
+ "model.layers.36.self_attn.v_proj": 1495,
+ "model.layers.36.self_attn.o_proj": 1495,
+ "model.layers.36.mlp.gate_proj": 2182,
+ "model.layers.36.mlp.up_proj": 2182,
+ "model.layers.36.mlp.down_proj": 2182,
+ "model.layers.37.self_attn.q_proj": 1495,
+ "model.layers.37.self_attn.k_proj": 1495,
+ "model.layers.37.self_attn.v_proj": 1495,
+ "model.layers.37.self_attn.o_proj": 1495,
+ "model.layers.37.mlp.gate_proj": 2182,
+ "model.layers.37.mlp.up_proj": 2182,
+ "model.layers.37.mlp.down_proj": 2182,
+ "model.layers.38.self_attn.q_proj": 1495,
+ "model.layers.38.self_attn.k_proj": 1495,
+ "model.layers.38.self_attn.v_proj": 1495,
+ "model.layers.38.self_attn.o_proj": 1495,
+ "model.layers.38.mlp.gate_proj": 2182,
+ "model.layers.38.mlp.up_proj": 2182,
+ "model.layers.38.mlp.down_proj": 2182,
+ "model.layers.39.self_attn.q_proj": 1495,
+ "model.layers.39.self_attn.k_proj": 1495,
+ "model.layers.39.self_attn.v_proj": 1495,
+ "model.layers.39.self_attn.o_proj": 1495,
+ "model.layers.39.mlp.gate_proj": 2182,
+ "model.layers.39.mlp.up_proj": 2182,
+ "model.layers.39.mlp.down_proj": 2182
+}

svd-llm-13b-unhealed/state_dict_fp16.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5625041bed874f21664b6a2bff3ce4722612302630028c9b60ac7f2cb8d7c605
+size 15475812809

svd-llm-13b/README.md

@@ -0,0 +1,29 @@
+---
+license: llama2
+base_model: NousResearch/Llama-2-13b-chat-hf
+tags: [svd, low-rank, model-compression, llama2, knowledge-distillation]
+---
+
+# SVD-LLM @ 7.74B (59.47% kept of LLaMA-2-13B-chat), KD-healed 60M tokens
+
+SVD-LLM (ICLR 2025): truncation-aware data whitening + closed-form parameter update, official code (@ 7538cca), uniform 41.58% removal across all 7 projection types of all 40 layers.
+
+- **HEALED checkpoint** (this dir). The matching **UNHEALED** pre-distillation weights are in
+  `svd-llm-13b-unhealed/` — do not mix them up: same `ranks.json`, different `state_dict_fp16.pt` (see
+  `provenance.json` SHA256 here vs the unhealed dir's record).
+- Healing: pure forward-KL distillation from frozen NousResearch/Llama-2-13b-chat-hf teacher,
+  60M tokens, held-out KL **2.0904 -> 0.3007**, probe MMLU (5-shot, limit 40) **24.56% -> 46.54%**.
+- Identical recipe/protocol to the TR-3 13B and the 7B SVD baselines (see their dirs).
+
+## Load
+
+```python
+import json
+from lowrank_linear import build, materialize_all
+model = build("NousResearch/Llama-2-13b-chat-hf", json.load(open("ranks.json")),
+              "state_dict_fp16.pt", device="cuda")
+materialize_all(model)   # inference-only fusion U@V
+```
+
+Full provenance (data mixture + revisions, decontamination, SHA256 of every input and
+checkpoint, 12 x 5M milestone trajectory) in `provenance.json`.

svd-llm-13b/config.json

@@ -0,0 +1,26 @@
+{
+  "_name_or_path": null,
+  "architectures": [
+    "LlamaForCausalLM"
+  ],
+  "bos_token_id": 1,
+  "eos_token_id": 2,
+  "hidden_act": "silu",
+  "hidden_size": 5120,
+  "initializer_range": 0.02,
+  "intermediate_size": 13824,
+  "max_position_embeddings": 4096,
+  "model_type": "llama",
+  "num_attention_heads": 40,
+  "num_hidden_layers": 40,
+  "num_key_value_heads": 40,
+  "pad_token_id": 0,
+  "pretraining_tp": 1,
+  "rms_norm_eps": 1e-05,
+  "rope_scaling": null,
+  "tie_word_embeddings": false,
+  "torch_dtype": "float16",
+  "transformers_version": "4.31.0.dev0",
+  "use_cache": true,
+  "vocab_size": 32000
+}

svd-llm-13b/lowrank_linear.py

@@ -0,0 +1,83 @@
+"""Minimal standalone loader for the SVD-baseline compressed LLaMA-2-7B checkpoints.
+
+No dependency beyond torch + transformers. The state_dict stores every compressed
+projection as two factors of a LowRankLinear (A: out x k, B: k x in); ranks.json
+holds the per-layer integer rank of each projection. Untouched parameters
+(embeddings, norms) keep their stock transformers names.
+
+Usage:
+    import json, torch
+    from lowrank_linear import build
+    model = build(
+        base_id="NousResearch/Llama-2-7b-chat-hf",
+        ranks=json.load(open("ranks.json")),
+        state_dict_path="state_dict_fp16.pt",
+        device="cuda", dtype=torch.float16,
+    )
+    tokenizer = AutoTokenizer.from_pretrained("NousResearch/Llama-2-7b-chat-hf")
+    out = model.generate(tokenizer("2+2=", return_tensors="pt").to("cuda").input_ids,
+                         max_new_tokens=16, do_sample=False)
+    print(tokenizer.decode(out[0]))
+"""
+
+from __future__ import annotations
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+
+class LowRankLinear(nn.Module):
+    """y = A (B x).  A: out x r, B: r x in.  (LLaMA projections carry no bias.)"""
+
+    def __init__(self, A: torch.Tensor, B: torch.Tensor):
+        super().__init__()
+        self.A = nn.Parameter(A)
+        self.B = nn.Parameter(B)
+
+    def forward(self, x):
+        return F.linear(F.linear(x, self.B), self.A)
+
+    @torch.no_grad()
+    def materialize_(self) -> nn.Linear:
+        """Replace this module with the dense W = A @ B (inference only; never call
+        while training -- it cuts the gradient path)."""
+        W = (self.A @ self.B).to(self.A.dtype)
+        dense = nn.Linear(self.B.shape[1], self.A.shape[0], bias=False,
+                          device=W.device, dtype=W.dtype)
+        dense.weight.data = W
+        return dense
+
+
+def build(base_id: str, ranks: dict, state_dict_path: str,
+          device: str = "cuda", dtype=torch.float16):
+    from transformers import AutoModelForCausalLM
+
+    model = AutoModelForCausalLM.from_pretrained(base_id, torch_dtype=dtype)
+    search = {}
+    for name, mod in model.named_modules():
+        if isinstance(mod, nn.Linear):
+            search[name] = mod
+    for path, k in ranks.items():
+        out_f, in_f = search[path].weight.shape
+        model_mod = model
+        for p in path.split(".")[:-1]:
+            model_mod = getattr(model_mod, p)
+        setattr(model_mod, path.split(".")[-1],
+                LowRankLinear(torch.zeros(out_f, int(k), dtype=dtype),
+                              torch.zeros(int(k), in_f, dtype=dtype)))
+    sd = torch.load(state_dict_path, map_location="cpu", weights_only=True)
+    model.load_state_dict(sd, strict=True)
+    del sd
+    return model.to(device)
+
+
+def materialize_all(model):
+    """Fuse every LowRankLinear back into a dense nn.Linear in place."""
+    for name, mod in list(model.named_modules()):
+        if isinstance(mod, LowRankLinear):
+            parent = model
+            for p in name.split(".")[:-1]:
+                parent = getattr(parent, p)
+            setattr(parent, name.split(".")[-1], mod.materialize_())
+    return model

svd-llm-13b/provenance.json

@@ -0,0 +1,199 @@
+{
+  "method": "KD healing: alpha*T^2*KL(teacher||student) + (1-alpha)*CE",
+  "student_init": "/data1/rkp00022/zixuan/projects/TR3_main/checkpoints/svdllm1_r04158_tr3fmt/state_dict_fp16.pt",
+  "teacher": "NousResearch/Llama-2-13b-chat-hf",
+  "alpha": 1.0,
+  "temperature": 1.0,
+  "lr": 2e-05,
+  "tokens_seen": 59998208,
+  "epochs": 0.228,
+  "global_batch_tokens": 32768,
+  "dataset_mixture_pre_exclusion": {
+    "alpaca": 50000,
+    "openhermes": 200000,
+    "ultrachat": 100000
+  },
+  "dataset_revisions": {
+    "tatsu-lab/alpaca": "dce01c9b08f87459cf36a430d809084718273017",
+    "teknium/OpenHermes-2.5": "b82037821055c377bed0d495e72e46de3bc72e84",
+    "HuggingFaceH4/ultrachat_200k": "8049631c405ae6576f93f445c6b8166f76f5505a"
+  },
+  "canonical_pool_excluded_texts": 10000,
+  "train_texts_after_exclusion": 356000,
+  "chat_texts": 340000,
+  "mmlu_aux_texts_kept": 16000,
+  "mmlu_aux_split": "cais/mmlu auxiliary_train (official TRAIN split, disjoint from the evaluated test split; audit verified 0/99842 exact overlap)",
+  "mmlu_aux_stats": {
+    "requested_groups": 16000,
+    "rendered_groups": 16000,
+    "shots": 5,
+    "rows_total": 99842,
+    "rows_dropped_by_decontam": 71,
+    "rows_clean": 99771,
+    "revision": "c30699e8356da336a370243923dbaf21066bb9fe",
+    "format": "MMLU-shaped (NOT byte-identical to lm-eval; auxiliary_train has no subject)"
+  },
+  "gold_label_supervision": false,
+  "attribution_note": "alpha=1.0 => pure KD, no gold-label CE, so any MMLU change is attributable to distillation on in-domain text. alpha<1 adds CE on text containing answers = labeled MCQ training, which is a DIFFERENT and weaker claim.",
+  "train_tokens_total": 263112704,
+  "decontamination": {
+    "removed": {
+      "alpaca": 11,
+      "openhermes": 120,
+      "ultrachat": 33
+    }
+  },
+  "start_kl": 2.0904040336608887,
+  "final_kl": 0.30065654776990414,
+  "start_mmlu": 0.24561403508771928,
+  "final_mmlu": 0.46535087719298246,
+  "history": [
+    {
+      "tokens_m": 5.01,
+      "step": 153,
+      "held_out_kl": 0.5967434793710709,
+      "mmlu": 0.3929824561403509
+    },
+    {
+      "tokens_m": 10.03,
+      "step": 306,
+      "held_out_kl": 0.5406020767986774,
+      "mmlu": 0.3956140350877193
+    },
+    {
+      "tokens_m": 15.01,
+      "step": 458,
+      "held_out_kl": 0.5206542909145355,
+      "mmlu": 0.4083333333333333
+    },
+    {
+      "tokens_m": 20.02,
+      "step": 611,
+      "held_out_kl": 0.4988052695989609,
+      "mmlu": 0.40087719298245617
+    },
+    {
+      "tokens_m": 25.0,
+      "step": 763,
+      "held_out_kl": 0.4575178399682045,
+      "mmlu": 0.40350877192982454
+    },
+    {
+      "tokens_m": 30.02,
+      "step": 916,
+      "held_out_kl": 0.4198819752782583,
+      "mmlu": 0.4298245614035088
+    },
+    {
+      "tokens_m": 35.03,
+      "step": 1069,
+      "held_out_kl": 0.38331497088074684,
+      "mmlu": 0.43157894736842106
+    },
+    {
+      "tokens_m": 40.01,
+      "step": 1221,
+      "held_out_kl": 0.34914770163595676,
+      "mmlu": 0.4425438596491228
+    },
+    {
+      "tokens_m": 45.02,
+      "step": 1374,
+      "held_out_kl": 0.32529739663004875,
+      "mmlu": 0.44956140350877194
+    },
+    {
+      "tokens_m": 50.0,
+      "step": 1526,
+      "held_out_kl": 0.31298027001321316,
+      "mmlu": 0.4587719298245614
+    },
+    {
+      "tokens_m": 55.02,
+      "step": 1679,
+      "held_out_kl": 0.30493565648794174,
+      "mmlu": 0.46403508771929824
+    },
+    {
+      "tokens_m": 60.0,
+      "step": 1831,
+      "held_out_kl": 0.30065654776990414,
+      "mmlu": 0.46535087719298246,
+      "final": true
+    }
+  ],
+  "teacher_mmlu_same_limit": 0.5442982456140351,
+  "checkpoint": "/data1/rkp00022/zixuan/projects/TR3_main/checkpoints/kdheal_svdllm1r04158_h1/state_dict_fp16.pt",
+  "sha256": "42b2f5c959e8fdd07ef526e6d182cab2f7685354859e1262b035eef39ba98387",
+  "seed": 1234,
+  "mmlu_limit_per_subtask": 40,
+  "teacher_revision": "d73f5fa9c4bc135502e04c27b39660747172d76b",
+  "student_init_sha256": "5625041bed874f21664b6a2bff3ce4722612302630028c9b60ac7f2cb8d7c605",
+  "ranks_file": "/data1/rkp00022/zixuan/projects/TR3_main/checkpoints/svdllm1_r04158_tr3fmt/ranks.json",
+  "ranks_sha256": "5c9ed0db91c9c81c51e62e6f82c7c2772478cc353c6f28040164843e64219b55",
+  "input_identity_gate": {
+    "student_ckpt": {
+      "key": "svdllm1_r04158_tr3fmt",
+      "sha256": "5625041bed874f21664b6a2bff3ce4722612302630028c9b60ac7f2cb8d7c605",
+      "status": "unregistered_allowed"
+    },
+    "ranks": {
+      "key": "svdllm1_r04158_tr3fmt",
+      "sha256": "5c9ed0db91c9c81c51e62e6f82c7c2772478cc353c6f28040164843e64219b55",
+      "status": "unregistered_allowed"
+    }
+  },
+  "train_seed": 1234,
+  "mmlu_seed": 1234,
+  "dynamic_batch": false,
+  "dbl": null,
+  "snapshot_milestones": true,
+  "snapshots": [
+    "(PosixPath('/data1/rkp00022/zixuan/projects/TR3_main/checkpoints/kdheal_svdllm1r04158_h1_snap15m'), 15.007744)",
+    "(PosixPath('/data1/rkp00022/zixuan/projects/TR3_main/checkpoints/kdheal_svdllm1r04158_h1_snap30m'), 30.015488)",
+    "(PosixPath('/data1/rkp00022/zixuan/projects/TR3_main/checkpoints/kdheal_svdllm1r04158_h1_snap45m'), 45.023232)",
+    "(PosixPath('/data1/rkp00022/zixuan/projects/TR3_main/checkpoints/kdheal_svdllm1r04158_h1_snap50m'), 50.003968)",
+    "(PosixPath('/data1/rkp00022/zixuan/projects/TR3_main/checkpoints/kdheal_svdllm1r04158_h1_snap55m'), 55.017472)"
+  ],
+  "stream_skip_tokens_m": 0.0,
+  "stream_blocks_consumed": [
+    0,
+    58592
+  ],
+  "stream_note": "the block order is a seeded permutation consumed from index 0; stages sharing a --seed with stream_skip=0 re-train on the SAME blocks. [start, end] block indices are recorded so any continuation can be checked for overlap with what earlier stages consumed.",
+  "student_master_dtype": "bf16",
+  "fake_quant": null,
+  "rebalance_report": null,
+  "student_param_count": 7737799680,
+  "micro": 8,
+  "grad_accum": 4,
+  "seq_len": 1024,
+  "held_out_source": "canonical: mixture 2000/4000/4000, seed 1234, seq 1024, blocks 256:264 (identical to the compression gates)",
+  "decontam_stats": {
+    "mmlu": {
+      "items": 14042,
+      "new_ngrams": 764261,
+      "revision": "c30699e8356da336a370243923dbaf21066bb9fe"
+    },
+    "hellaswag": {
+      "items": 10042,
+      "new_ngrams": 1334302,
+      "revision": "218ec52e09a7e7462a5400043bb9a69a41d06b76"
+    },
+    "boolq": {
+      "items": 3270,
+      "new_ngrams": 276806,
+      "revision": "3de24cf8022e94f4ee4b9d55a6f539891524d646"
+    },
+    "triviaqa": {
+      "items": 17944,
+      "new_ngrams": 32608,
+      "revision": "0f7faf33a3908546c6fd5b73a660e0f8ff173c2f"
+    },
+    "gsm8k": {
+      "items": 1319,
+      "new_ngrams": 46281,
+      "revision": "740312add88f781978c0658806c59bc2815b9866"
+    }
+  }
+}

svd-llm-13b/ranks.json

@@ -0,0 +1,282 @@
+{
+ "model.layers.0.self_attn.q_proj": 1495,
+ "model.layers.0.self_attn.k_proj": 1495,
+ "model.layers.0.self_attn.v_proj": 1495,
+ "model.layers.0.self_attn.o_proj": 1495,
+ "model.layers.0.mlp.gate_proj": 2182,
+ "model.layers.0.mlp.up_proj": 2182,
+ "model.layers.0.mlp.down_proj": 2182,
+ "model.layers.1.self_attn.q_proj": 1495,
+ "model.layers.1.self_attn.k_proj": 1495,
+ "model.layers.1.self_attn.v_proj": 1495,
+ "model.layers.1.self_attn.o_proj": 1495,
+ "model.layers.1.mlp.gate_proj": 2182,
+ "model.layers.1.mlp.up_proj": 2182,
+ "model.layers.1.mlp.down_proj": 2182,
+ "model.layers.2.self_attn.q_proj": 1495,
+ "model.layers.2.self_attn.k_proj": 1495,
+ "model.layers.2.self_attn.v_proj": 1495,
+ "model.layers.2.self_attn.o_proj": 1495,
+ "model.layers.2.mlp.gate_proj": 2182,
+ "model.layers.2.mlp.up_proj": 2182,
+ "model.layers.2.mlp.down_proj": 2182,
+ "model.layers.3.self_attn.q_proj": 1495,
+ "model.layers.3.self_attn.k_proj": 1495,
+ "model.layers.3.self_attn.v_proj": 1495,
+ "model.layers.3.self_attn.o_proj": 1495,
+ "model.layers.3.mlp.gate_proj": 2182,
+ "model.layers.3.mlp.up_proj": 2182,
+ "model.layers.3.mlp.down_proj": 2182,
+ "model.layers.4.self_attn.q_proj": 1495,
+ "model.layers.4.self_attn.k_proj": 1495,
+ "model.layers.4.self_attn.v_proj": 1495,
+ "model.layers.4.self_attn.o_proj": 1495,
+ "model.layers.4.mlp.gate_proj": 2182,
+ "model.layers.4.mlp.up_proj": 2182,
+ "model.layers.4.mlp.down_proj": 2182,
+ "model.layers.5.self_attn.q_proj": 1495,
+ "model.layers.5.self_attn.k_proj": 1495,
+ "model.layers.5.self_attn.v_proj": 1495,
+ "model.layers.5.self_attn.o_proj": 1495,
+ "model.layers.5.mlp.gate_proj": 2182,
+ "model.layers.5.mlp.up_proj": 2182,
+ "model.layers.5.mlp.down_proj": 2182,
+ "model.layers.6.self_attn.q_proj": 1495,
+ "model.layers.6.self_attn.k_proj": 1495,
+ "model.layers.6.self_attn.v_proj": 1495,
+ "model.layers.6.self_attn.o_proj": 1495,
+ "model.layers.6.mlp.gate_proj": 2182,
+ "model.layers.6.mlp.up_proj": 2182,
+ "model.layers.6.mlp.down_proj": 2182,
+ "model.layers.7.self_attn.q_proj": 1495,
+ "model.layers.7.self_attn.k_proj": 1495,
+ "model.layers.7.self_attn.v_proj": 1495,
+ "model.layers.7.self_attn.o_proj": 1495,
+ "model.layers.7.mlp.gate_proj": 2182,
+ "model.layers.7.mlp.up_proj": 2182,
+ "model.layers.7.mlp.down_proj": 2182,
+ "model.layers.8.self_attn.q_proj": 1495,
+ "model.layers.8.self_attn.k_proj": 1495,
+ "model.layers.8.self_attn.v_proj": 1495,
+ "model.layers.8.self_attn.o_proj": 1495,
+ "model.layers.8.mlp.gate_proj": 2182,
+ "model.layers.8.mlp.up_proj": 2182,
+ "model.layers.8.mlp.down_proj": 2182,
+ "model.layers.9.self_attn.q_proj": 1495,
+ "model.layers.9.self_attn.k_proj": 1495,
+ "model.layers.9.self_attn.v_proj": 1495,
+ "model.layers.9.self_attn.o_proj": 1495,
+ "model.layers.9.mlp.gate_proj": 2182,
+ "model.layers.9.mlp.up_proj": 2182,
+ "model.layers.9.mlp.down_proj": 2182,
+ "model.layers.10.self_attn.q_proj": 1495,
+ "model.layers.10.self_attn.k_proj": 1495,
+ "model.layers.10.self_attn.v_proj": 1495,
+ "model.layers.10.self_attn.o_proj": 1495,
+ "model.layers.10.mlp.gate_proj": 2182,
+ "model.layers.10.mlp.up_proj": 2182,
+ "model.layers.10.mlp.down_proj": 2182,
+ "model.layers.11.self_attn.q_proj": 1495,
+ "model.layers.11.self_attn.k_proj": 1495,
+ "model.layers.11.self_attn.v_proj": 1495,
+ "model.layers.11.self_attn.o_proj": 1495,
+ "model.layers.11.mlp.gate_proj": 2182,
+ "model.layers.11.mlp.up_proj": 2182,
+ "model.layers.11.mlp.down_proj": 2182,
+ "model.layers.12.self_attn.q_proj": 1495,
+ "model.layers.12.self_attn.k_proj": 1495,
+ "model.layers.12.self_attn.v_proj": 1495,
+ "model.layers.12.self_attn.o_proj": 1495,
+ "model.layers.12.mlp.gate_proj": 2182,
+ "model.layers.12.mlp.up_proj": 2182,
+ "model.layers.12.mlp.down_proj": 2182,
+ "model.layers.13.self_attn.q_proj": 1495,
+ "model.layers.13.self_attn.k_proj": 1495,
+ "model.layers.13.self_attn.v_proj": 1495,
+ "model.layers.13.self_attn.o_proj": 1495,
+ "model.layers.13.mlp.gate_proj": 2182,
+ "model.layers.13.mlp.up_proj": 2182,
+ "model.layers.13.mlp.down_proj": 2182,
+ "model.layers.14.self_attn.q_proj": 1495,
+ "model.layers.14.self_attn.k_proj": 1495,
+ "model.layers.14.self_attn.v_proj": 1495,
+ "model.layers.14.self_attn.o_proj": 1495,
+ "model.layers.14.mlp.gate_proj": 2182,
+ "model.layers.14.mlp.up_proj": 2182,
+ "model.layers.14.mlp.down_proj": 2182,
+ "model.layers.15.self_attn.q_proj": 1495,
+ "model.layers.15.self_attn.k_proj": 1495,
+ "model.layers.15.self_attn.v_proj": 1495,
+ "model.layers.15.self_attn.o_proj": 1495,
+ "model.layers.15.mlp.gate_proj": 2182,
+ "model.layers.15.mlp.up_proj": 2182,
+ "model.layers.15.mlp.down_proj": 2182,
+ "model.layers.16.self_attn.q_proj": 1495,
+ "model.layers.16.self_attn.k_proj": 1495,
+ "model.layers.16.self_attn.v_proj": 1495,
+ "model.layers.16.self_attn.o_proj": 1495,
+ "model.layers.16.mlp.gate_proj": 2182,
+ "model.layers.16.mlp.up_proj": 2182,
+ "model.layers.16.mlp.down_proj": 2182,
+ "model.layers.17.self_attn.q_proj": 1495,
+ "model.layers.17.self_attn.k_proj": 1495,
+ "model.layers.17.self_attn.v_proj": 1495,
+ "model.layers.17.self_attn.o_proj": 1495,
+ "model.layers.17.mlp.gate_proj": 2182,
+ "model.layers.17.mlp.up_proj": 2182,
+ "model.layers.17.mlp.down_proj": 2182,
+ "model.layers.18.self_attn.q_proj": 1495,
+ "model.layers.18.self_attn.k_proj": 1495,
+ "model.layers.18.self_attn.v_proj": 1495,
+ "model.layers.18.self_attn.o_proj": 1495,
+ "model.layers.18.mlp.gate_proj": 2182,
+ "model.layers.18.mlp.up_proj": 2182,
+ "model.layers.18.mlp.down_proj": 2182,
+ "model.layers.19.self_attn.q_proj": 1495,
+ "model.layers.19.self_attn.k_proj": 1495,
+ "model.layers.19.self_attn.v_proj": 1495,
+ "model.layers.19.self_attn.o_proj": 1495,
+ "model.layers.19.mlp.gate_proj": 2182,
+ "model.layers.19.mlp.up_proj": 2182,
+ "model.layers.19.mlp.down_proj": 2182,
+ "model.layers.20.self_attn.q_proj": 1495,
+ "model.layers.20.self_attn.k_proj": 1495,
+ "model.layers.20.self_attn.v_proj": 1495,
+ "model.layers.20.self_attn.o_proj": 1495,
+ "model.layers.20.mlp.gate_proj": 2182,
+ "model.layers.20.mlp.up_proj": 2182,
+ "model.layers.20.mlp.down_proj": 2182,
+ "model.layers.21.self_attn.q_proj": 1495,
+ "model.layers.21.self_attn.k_proj": 1495,
+ "model.layers.21.self_attn.v_proj": 1495,
+ "model.layers.21.self_attn.o_proj": 1495,
+ "model.layers.21.mlp.gate_proj": 2182,
+ "model.layers.21.mlp.up_proj": 2182,
+ "model.layers.21.mlp.down_proj": 2182,
+ "model.layers.22.self_attn.q_proj": 1495,
+ "model.layers.22.self_attn.k_proj": 1495,
+ "model.layers.22.self_attn.v_proj": 1495,
+ "model.layers.22.self_attn.o_proj": 1495,
+ "model.layers.22.mlp.gate_proj": 2182,
+ "model.layers.22.mlp.up_proj": 2182,
+ "model.layers.22.mlp.down_proj": 2182,
+ "model.layers.23.self_attn.q_proj": 1495,
+ "model.layers.23.self_attn.k_proj": 1495,
+ "model.layers.23.self_attn.v_proj": 1495,
+ "model.layers.23.self_attn.o_proj": 1495,
+ "model.layers.23.mlp.gate_proj": 2182,
+ "model.layers.23.mlp.up_proj": 2182,
+ "model.layers.23.mlp.down_proj": 2182,
+ "model.layers.24.self_attn.q_proj": 1495,
+ "model.layers.24.self_attn.k_proj": 1495,
+ "model.layers.24.self_attn.v_proj": 1495,
+ "model.layers.24.self_attn.o_proj": 1495,
+ "model.layers.24.mlp.gate_proj": 2182,
+ "model.layers.24.mlp.up_proj": 2182,
+ "model.layers.24.mlp.down_proj": 2182,
+ "model.layers.25.self_attn.q_proj": 1495,
+ "model.layers.25.self_attn.k_proj": 1495,
+ "model.layers.25.self_attn.v_proj": 1495,
+ "model.layers.25.self_attn.o_proj": 1495,
+ "model.layers.25.mlp.gate_proj": 2182,
+ "model.layers.25.mlp.up_proj": 2182,
+ "model.layers.25.mlp.down_proj": 2182,
+ "model.layers.26.self_attn.q_proj": 1495,
+ "model.layers.26.self_attn.k_proj": 1495,
+ "model.layers.26.self_attn.v_proj": 1495,
+ "model.layers.26.self_attn.o_proj": 1495,
+ "model.layers.26.mlp.gate_proj": 2182,
+ "model.layers.26.mlp.up_proj": 2182,
+ "model.layers.26.mlp.down_proj": 2182,
+ "model.layers.27.self_attn.q_proj": 1495,
+ "model.layers.27.self_attn.k_proj": 1495,
+ "model.layers.27.self_attn.v_proj": 1495,
+ "model.layers.27.self_attn.o_proj": 1495,
+ "model.layers.27.mlp.gate_proj": 2182,
+ "model.layers.27.mlp.up_proj": 2182,
+ "model.layers.27.mlp.down_proj": 2182,
+ "model.layers.28.self_attn.q_proj": 1495,
+ "model.layers.28.self_attn.k_proj": 1495,
+ "model.layers.28.self_attn.v_proj": 1495,
+ "model.layers.28.self_attn.o_proj": 1495,
+ "model.layers.28.mlp.gate_proj": 2182,
+ "model.layers.28.mlp.up_proj": 2182,
+ "model.layers.28.mlp.down_proj": 2182,
+ "model.layers.29.self_attn.q_proj": 1495,
+ "model.layers.29.self_attn.k_proj": 1495,
+ "model.layers.29.self_attn.v_proj": 1495,
+ "model.layers.29.self_attn.o_proj": 1495,
+ "model.layers.29.mlp.gate_proj": 2182,
+ "model.layers.29.mlp.up_proj": 2182,
+ "model.layers.29.mlp.down_proj": 2182,
+ "model.layers.30.self_attn.q_proj": 1495,
+ "model.layers.30.self_attn.k_proj": 1495,
+ "model.layers.30.self_attn.v_proj": 1495,
+ "model.layers.30.self_attn.o_proj": 1495,
+ "model.layers.30.mlp.gate_proj": 2182,
+ "model.layers.30.mlp.up_proj": 2182,
+ "model.layers.30.mlp.down_proj": 2182,
+ "model.layers.31.self_attn.q_proj": 1495,
+ "model.layers.31.self_attn.k_proj": 1495,
+ "model.layers.31.self_attn.v_proj": 1495,
+ "model.layers.31.self_attn.o_proj": 1495,
+ "model.layers.31.mlp.gate_proj": 2182,
+ "model.layers.31.mlp.up_proj": 2182,
+ "model.layers.31.mlp.down_proj": 2182,
+ "model.layers.32.self_attn.q_proj": 1495,
+ "model.layers.32.self_attn.k_proj": 1495,
+ "model.layers.32.self_attn.v_proj": 1495,
+ "model.layers.32.self_attn.o_proj": 1495,
+ "model.layers.32.mlp.gate_proj": 2182,
+ "model.layers.32.mlp.up_proj": 2182,
+ "model.layers.32.mlp.down_proj": 2182,
+ "model.layers.33.self_attn.q_proj": 1495,
+ "model.layers.33.self_attn.k_proj": 1495,
+ "model.layers.33.self_attn.v_proj": 1495,
+ "model.layers.33.self_attn.o_proj": 1495,
+ "model.layers.33.mlp.gate_proj": 2182,
+ "model.layers.33.mlp.up_proj": 2182,
+ "model.layers.33.mlp.down_proj": 2182,
+ "model.layers.34.self_attn.q_proj": 1495,
+ "model.layers.34.self_attn.k_proj": 1495,
+ "model.layers.34.self_attn.v_proj": 1495,
+ "model.layers.34.self_attn.o_proj": 1495,
+ "model.layers.34.mlp.gate_proj": 2182,
+ "model.layers.34.mlp.up_proj": 2182,
+ "model.layers.34.mlp.down_proj": 2182,
+ "model.layers.35.self_attn.q_proj": 1495,
+ "model.layers.35.self_attn.k_proj": 1495,
+ "model.layers.35.self_attn.v_proj": 1495,
+ "model.layers.35.self_attn.o_proj": 1495,
+ "model.layers.35.mlp.gate_proj": 2182,
+ "model.layers.35.mlp.up_proj": 2182,
+ "model.layers.35.mlp.down_proj": 2182,
+ "model.layers.36.self_attn.q_proj": 1495,
+ "model.layers.36.self_attn.k_proj": 1495,
+ "model.layers.36.self_attn.v_proj": 1495,
+ "model.layers.36.self_attn.o_proj": 1495,
+ "model.layers.36.mlp.gate_proj": 2182,
+ "model.layers.36.mlp.up_proj": 2182,
+ "model.layers.36.mlp.down_proj": 2182,
+ "model.layers.37.self_attn.q_proj": 1495,
+ "model.layers.37.self_attn.k_proj": 1495,
+ "model.layers.37.self_attn.v_proj": 1495,
+ "model.layers.37.self_attn.o_proj": 1495,
+ "model.layers.37.mlp.gate_proj": 2182,
+ "model.layers.37.mlp.up_proj": 2182,
+ "model.layers.37.mlp.down_proj": 2182,
+ "model.layers.38.self_attn.q_proj": 1495,
+ "model.layers.38.self_attn.k_proj": 1495,
+ "model.layers.38.self_attn.v_proj": 1495,
+ "model.layers.38.self_attn.o_proj": 1495,
+ "model.layers.38.mlp.gate_proj": 2182,
+ "model.layers.38.mlp.up_proj": 2182,
+ "model.layers.38.mlp.down_proj": 2182,
+ "model.layers.39.self_attn.q_proj": 1495,
+ "model.layers.39.self_attn.k_proj": 1495,
+ "model.layers.39.self_attn.v_proj": 1495,
+ "model.layers.39.self_attn.o_proj": 1495,
+ "model.layers.39.mlp.gate_proj": 2182,
+ "model.layers.39.mlp.up_proj": 2182,
+ "model.layers.39.mlp.down_proj": 2182
+}

svd-llm-13b/state_dict_fp16.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:42b2f5c959e8fdd07ef526e6d182cab2f7685354859e1262b035eef39ba98387
+size 15475814672

svd-llm-v2-13b-unhealed/README.md

@@ -0,0 +1,16 @@
+---
+license: llama2
+base_model: NousResearch/Llama-2-13b-chat-hf
+tags: [svd, low-rank, model-compression, llama2]
+---
+
+# SVD-LLM V2 @ 7.74B (59.47% kept of LLaMA-2-13B-chat), UNHEALED
+
+SVD-LLM V2 (NAACL 2025): heterogeneous ratio allocation (Alg. 1) + loss-optimized truncation (Alg. 2). Official code never released; our faithful implementation (see per-layer ranks in ranks.json). Exact budget match: 7,740,286,464 vs target 7,740,282,880.
+
+- **UNHEALED checkpoint**: raw post-training compression weights, NO distillation. The KD-healed
+  counterpart is in `svd-llm-v2-13b/`. This dir exists so the healing delta is
+  measurable; use it as the compression-stage artifact, not as a deployable model.
+- Same `ranks.json` as the healed dir; weights differ (see `provenance.json` SHA256).
+
+Load exactly as in the sibling dir (`lowrank_linear.py`).

svd-llm-v2-13b-unhealed/config.json

@@ -0,0 +1,26 @@
+{
+  "_name_or_path": null,
+  "architectures": [
+    "LlamaForCausalLM"
+  ],
+  "bos_token_id": 1,
+  "eos_token_id": 2,
+  "hidden_act": "silu",
+  "hidden_size": 5120,
+  "initializer_range": 0.02,
+  "intermediate_size": 13824,
+  "max_position_embeddings": 4096,
+  "model_type": "llama",
+  "num_attention_heads": 40,
+  "num_hidden_layers": 40,
+  "num_key_value_heads": 40,
+  "pad_token_id": 0,
+  "pretraining_tp": 1,
+  "rms_norm_eps": 1e-05,
+  "rope_scaling": null,
+  "tie_word_embeddings": false,
+  "torch_dtype": "float16",
+  "transformers_version": "4.31.0.dev0",
+  "use_cache": true,
+  "vocab_size": 32000
+}

svd-llm-v2-13b-unhealed/lowrank_linear.py

@@ -0,0 +1,83 @@
+"""Minimal standalone loader for the SVD-baseline compressed LLaMA-2-7B checkpoints.
+
+No dependency beyond torch + transformers. The state_dict stores every compressed
+projection as two factors of a LowRankLinear (A: out x k, B: k x in); ranks.json
+holds the per-layer integer rank of each projection. Untouched parameters
+(embeddings, norms) keep their stock transformers names.
+
+Usage:
+    import json, torch
+    from lowrank_linear import build
+    model = build(
+        base_id="NousResearch/Llama-2-7b-chat-hf",
+        ranks=json.load(open("ranks.json")),
+        state_dict_path="state_dict_fp16.pt",
+        device="cuda", dtype=torch.float16,
+    )
+    tokenizer = AutoTokenizer.from_pretrained("NousResearch/Llama-2-7b-chat-hf")
+    out = model.generate(tokenizer("2+2=", return_tensors="pt").to("cuda").input_ids,
+                         max_new_tokens=16, do_sample=False)
+    print(tokenizer.decode(out[0]))
+"""
+
+from __future__ import annotations
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+
+class LowRankLinear(nn.Module):
+    """y = A (B x).  A: out x r, B: r x in.  (LLaMA projections carry no bias.)"""
+
+    def __init__(self, A: torch.Tensor, B: torch.Tensor):
+        super().__init__()
+        self.A = nn.Parameter(A)
+        self.B = nn.Parameter(B)
+
+    def forward(self, x):
+        return F.linear(F.linear(x, self.B), self.A)
+
+    @torch.no_grad()
+    def materialize_(self) -> nn.Linear:
+        """Replace this module with the dense W = A @ B (inference only; never call
+        while training -- it cuts the gradient path)."""
+        W = (self.A @ self.B).to(self.A.dtype)
+        dense = nn.Linear(self.B.shape[1], self.A.shape[0], bias=False,
+                          device=W.device, dtype=W.dtype)
+        dense.weight.data = W
+        return dense
+
+
+def build(base_id: str, ranks: dict, state_dict_path: str,
+          device: str = "cuda", dtype=torch.float16):
+    from transformers import AutoModelForCausalLM
+
+    model = AutoModelForCausalLM.from_pretrained(base_id, torch_dtype=dtype)
+    search = {}
+    for name, mod in model.named_modules():
+        if isinstance(mod, nn.Linear):
+            search[name] = mod
+    for path, k in ranks.items():
+        out_f, in_f = search[path].weight.shape
+        model_mod = model
+        for p in path.split(".")[:-1]:
+            model_mod = getattr(model_mod, p)
+        setattr(model_mod, path.split(".")[-1],
+                LowRankLinear(torch.zeros(out_f, int(k), dtype=dtype),
+                              torch.zeros(int(k), in_f, dtype=dtype)))
+    sd = torch.load(state_dict_path, map_location="cpu", weights_only=True)
+    model.load_state_dict(sd, strict=True)
+    del sd
+    return model.to(device)
+
+
+def materialize_all(model):
+    """Fuse every LowRankLinear back into a dense nn.Linear in place."""
+    for name, mod in list(model.named_modules()):
+        if isinstance(mod, LowRankLinear):
+            parent = model
+            for p in name.split(".")[:-1]:
+                parent = getattr(parent, p)
+            setattr(parent, name.split(".")[-1], mod.materialize_())
+    return model

svd-llm-v2-13b-unhealed/provenance.json

@@ -0,0 +1,7 @@
+{
+ "kind": "unhealed SVD compression (TR3-format conversion, verified pure re-keying)",
+ "source_ckpt": "/data1/rkp00022/zixuan/projects/TR3_main/checkpoints/svdllm2_llama2_13b_chat_r04158/NousResearch_Llama_2_13b_chat_hf_svdllmv2_keep_0.5740.pt",
+ "converted_state_dict": "/data1/rkp00022/zixuan/projects/TR3_main/scratch/hf_push_13b/svd-llm-v2-13b-unhealed/state_dict_fp16.pt",
+ "state_dict_sha256": "f0dbe9c89696d52b78344dc7564c8c5bc6bb57ecf7cd95edaa26e6738db87e23",
+ "healed_counterpart": "svd-llm-v2-13b/state_dict_fp16.pt"
+}

svd-llm-v2-13b-unhealed/ranks.json

@@ -0,0 +1,282 @@
+{
+ "model.layers.0.self_attn.q_proj": 2560,
+ "model.layers.0.self_attn.k_proj": 2560,
+ "model.layers.0.self_attn.v_proj": 2560,
+ "model.layers.0.self_attn.o_proj": 2560,
+ "model.layers.0.mlp.gate_proj": 3237,
+ "model.layers.0.mlp.up_proj": 3231,
+ "model.layers.0.mlp.down_proj": 3674,
+ "model.layers.1.self_attn.q_proj": 2514,
+ "model.layers.1.self_attn.k_proj": 2553,
+ "model.layers.1.self_attn.v_proj": 2288,
+ "model.layers.1.self_attn.o_proj": 2560,
+ "model.layers.1.mlp.gate_proj": 2862,
+ "model.layers.1.mlp.up_proj": 2816,
+ "model.layers.1.mlp.down_proj": 2909,
+ "model.layers.2.self_attn.q_proj": 1842,
+ "model.layers.2.self_attn.k_proj": 1847,
+ "model.layers.2.self_attn.v_proj": 1839,
+ "model.layers.2.self_attn.o_proj": 2126,
+ "model.layers.2.mlp.gate_proj": 2506,
+ "model.layers.2.mlp.up_proj": 2500,
+ "model.layers.2.mlp.down_proj": 2665,
+ "model.layers.3.self_attn.q_proj": 1735,
+ "model.layers.3.self_attn.k_proj": 1735,
+ "model.layers.3.self_attn.v_proj": 1742,
+ "model.layers.3.self_attn.o_proj": 2020,
+ "model.layers.3.mlp.gate_proj": 2363,
+ "model.layers.3.mlp.up_proj": 2369,
+ "model.layers.3.mlp.down_proj": 2569,
+ "model.layers.4.self_attn.q_proj": 1618,
+ "model.layers.4.self_attn.k_proj": 1618,
+ "model.layers.4.self_attn.v_proj": 1651,
+ "model.layers.4.self_attn.o_proj": 1888,
+ "model.layers.4.mlp.gate_proj": 2302,
+ "model.layers.4.mlp.up_proj": 2310,
+ "model.layers.4.mlp.down_proj": 2471,
+ "model.layers.5.self_attn.q_proj": 1603,
+ "model.layers.5.self_attn.k_proj": 1601,
+ "model.layers.5.self_attn.v_proj": 1624,
+ "model.layers.5.self_attn.o_proj": 1818,
+ "model.layers.5.mlp.gate_proj": 2253,
+ "model.layers.5.mlp.up_proj": 2262,
+ "model.layers.5.mlp.down_proj": 2388,
+ "model.layers.6.self_attn.q_proj": 1555,
+ "model.layers.6.self_attn.k_proj": 1551,
+ "model.layers.6.self_attn.v_proj": 1588,
+ "model.layers.6.self_attn.o_proj": 1778,
+ "model.layers.6.mlp.gate_proj": 2218,
+ "model.layers.6.mlp.up_proj": 2227,
+ "model.layers.6.mlp.down_proj": 2327,
+ "model.layers.7.self_attn.q_proj": 1556,
+ "model.layers.7.self_attn.k_proj": 1552,
+ "model.layers.7.self_attn.v_proj": 1590,
+ "model.layers.7.self_attn.o_proj": 1701,
+ "model.layers.7.mlp.gate_proj": 2199,
+ "model.layers.7.mlp.up_proj": 2208,
+ "model.layers.7.mlp.down_proj": 2277,
+ "model.layers.8.self_attn.q_proj": 1535,
+ "model.layers.8.self_attn.k_proj": 1531,
+ "model.layers.8.self_attn.v_proj": 1565,
+ "model.layers.8.self_attn.o_proj": 1663,
+ "model.layers.8.mlp.gate_proj": 2184,
+ "model.layers.8.mlp.up_proj": 2190,
+ "model.layers.8.mlp.down_proj": 2244,
+ "model.layers.9.self_attn.q_proj": 1504,
+ "model.layers.9.self_attn.k_proj": 1501,
+ "model.layers.9.self_attn.v_proj": 1544,
+ "model.layers.9.self_attn.o_proj": 1630,
+ "model.layers.9.mlp.gate_proj": 2169,
+ "model.layers.9.mlp.up_proj": 2170,
+ "model.layers.9.mlp.down_proj": 2219,
+ "model.layers.10.self_attn.q_proj": 1507,
+ "model.layers.10.self_attn.k_proj": 1506,
+ "model.layers.10.self_attn.v_proj": 1551,
+ "model.layers.10.self_attn.o_proj": 1583,
+ "model.layers.10.mlp.gate_proj": 2166,
+ "model.layers.10.mlp.up_proj": 2164,
+ "model.layers.10.mlp.down_proj": 2197,
+ "model.layers.11.self_attn.q_proj": 1509,
+ "model.layers.11.self_attn.k_proj": 1503,
+ "model.layers.11.self_attn.v_proj": 1554,
+ "model.layers.11.self_attn.o_proj": 1577,
+ "model.layers.11.mlp.gate_proj": 2165,
+ "model.layers.11.mlp.up_proj": 2159,
+ "model.layers.11.mlp.down_proj": 2180,
+ "model.layers.12.self_attn.q_proj": 1501,
+ "model.layers.12.self_attn.k_proj": 1495,
+ "model.layers.12.self_attn.v_proj": 1534,
+ "model.layers.12.self_attn.o_proj": 1552,
+ "model.layers.12.mlp.gate_proj": 2158,
+ "model.layers.12.mlp.up_proj": 2148,
+ "model.layers.12.mlp.down_proj": 2164,
+ "model.layers.13.self_attn.q_proj": 1494,
+ "model.layers.13.self_attn.k_proj": 1494,
+ "model.layers.13.self_attn.v_proj": 1514,
+ "model.layers.13.self_attn.o_proj": 1534,
+ "model.layers.13.mlp.gate_proj": 2149,
+ "model.layers.13.mlp.up_proj": 2138,
+ "model.layers.13.mlp.down_proj": 2142,
+ "model.layers.14.self_attn.q_proj": 1482,
+ "model.layers.14.self_attn.k_proj": 1480,
+ "model.layers.14.self_attn.v_proj": 1512,
+ "model.layers.14.self_attn.o_proj": 1499,
+ "model.layers.14.mlp.gate_proj": 2141,
+ "model.layers.14.mlp.up_proj": 2129,
+ "model.layers.14.mlp.down_proj": 2122,
+ "model.layers.15.self_attn.q_proj": 1486,
+ "model.layers.15.self_attn.k_proj": 1482,
+ "model.layers.15.self_attn.v_proj": 1507,
+ "model.layers.15.self_attn.o_proj": 1512,
+ "model.layers.15.mlp.gate_proj": 2131,
+ "model.layers.15.mlp.up_proj": 2117,
+ "model.layers.15.mlp.down_proj": 2103,
+ "model.layers.16.self_attn.q_proj": 1476,
+ "model.layers.16.self_attn.k_proj": 1471,
+ "model.layers.16.self_attn.v_proj": 1490,
+ "model.layers.16.self_attn.o_proj": 1504,
+ "model.layers.16.mlp.gate_proj": 2121,
+ "model.layers.16.mlp.up_proj": 2107,
+ "model.layers.16.mlp.down_proj": 2077,
+ "model.layers.17.self_attn.q_proj": 1468,
+ "model.layers.17.self_attn.k_proj": 1466,
+ "model.layers.17.self_attn.v_proj": 1482,
+ "model.layers.17.self_attn.o_proj": 1491,
+ "model.layers.17.mlp.gate_proj": 2108,
+ "model.layers.17.mlp.up_proj": 2098,
+ "model.layers.17.mlp.down_proj": 2066,
+ "model.layers.18.self_attn.q_proj": 1455,
+ "model.layers.18.self_attn.k_proj": 1453,
+ "model.layers.18.self_attn.v_proj": 1469,
+ "model.layers.18.self_attn.o_proj": 1498,
+ "model.layers.18.mlp.gate_proj": 2094,
+ "model.layers.18.mlp.up_proj": 2088,
+ "model.layers.18.mlp.down_proj": 2058,
+ "model.layers.19.self_attn.q_proj": 1461,
+ "model.layers.19.self_attn.k_proj": 1460,
+ "model.layers.19.self_attn.v_proj": 1468,
+ "model.layers.19.self_attn.o_proj": 1499,
+ "model.layers.19.mlp.gate_proj": 2083,
+ "model.layers.19.mlp.up_proj": 2079,
+ "model.layers.19.mlp.down_proj": 2044,
+ "model.layers.20.self_attn.q_proj": 1462,
+ "model.layers.20.self_attn.k_proj": 1462,
+ "model.layers.20.self_attn.v_proj": 1474,
+ "model.layers.20.self_attn.o_proj": 1480,
+ "model.layers.20.mlp.gate_proj": 2077,
+ "model.layers.20.mlp.up_proj": 2074,
+ "model.layers.20.mlp.down_proj": 2030,
+ "model.layers.21.self_attn.q_proj": 1464,
+ "model.layers.21.self_attn.k_proj": 1464,
+ "model.layers.21.self_attn.v_proj": 1463,
+ "model.layers.21.self_attn.o_proj": 1460,
+ "model.layers.21.mlp.gate_proj": 2072,
+ "model.layers.21.mlp.up_proj": 2072,
+ "model.layers.21.mlp.down_proj": 2010,
+ "model.layers.22.self_attn.q_proj": 1436,
+ "model.layers.22.self_attn.k_proj": 1437,
+ "model.layers.22.self_attn.v_proj": 1442,
+ "model.layers.22.self_attn.o_proj": 1491,
+ "model.layers.22.mlp.gate_proj": 2061,
+ "model.layers.22.mlp.up_proj": 2063,
+ "model.layers.22.mlp.down_proj": 2003,
+ "model.layers.23.self_attn.q_proj": 1442,
+ "model.layers.23.self_attn.k_proj": 1443,
+ "model.layers.23.self_attn.v_proj": 1442,
+ "model.layers.23.self_attn.o_proj": 1488,
+ "model.layers.23.mlp.gate_proj": 2051,
+ "model.layers.23.mlp.up_proj": 2056,
+ "model.layers.23.mlp.down_proj": 1994,
+ "model.layers.24.self_attn.q_proj": 1433,
+ "model.layers.24.self_attn.k_proj": 1435,
+ "model.layers.24.self_attn.v_proj": 1436,
+ "model.layers.24.self_attn.o_proj": 1495,
+ "model.layers.24.mlp.gate_proj": 2044,
+ "model.layers.24.mlp.up_proj": 2050,
+ "model.layers.24.mlp.down_proj": 1988,
+ "model.layers.25.self_attn.q_proj": 1428,
+ "model.layers.25.self_attn.k_proj": 1428,
+ "model.layers.25.self_attn.v_proj": 1425,
+ "model.layers.25.self_attn.o_proj": 1475,
+ "model.layers.25.mlp.gate_proj": 2035,
+ "model.layers.25.mlp.up_proj": 2042,
+ "model.layers.25.mlp.down_proj": 1978,
+ "model.layers.26.self_attn.q_proj": 1424,
+ "model.layers.26.self_attn.k_proj": 1424,
+ "model.layers.26.self_attn.v_proj": 1421,
+ "model.layers.26.self_attn.o_proj": 1474,
+ "model.layers.26.mlp.gate_proj": 2027,
+ "model.layers.26.mlp.up_proj": 2035,
+ "model.layers.26.mlp.down_proj": 1969,
+ "model.layers.27.self_attn.q_proj": 1408,
+ "model.layers.27.self_attn.k_proj": 1409,
+ "model.layers.27.self_attn.v_proj": 1410,
+ "model.layers.27.self_attn.o_proj": 1474,
+ "model.layers.27.mlp.gate_proj": 2019,
+ "model.layers.27.mlp.up_proj": 2027,
+ "model.layers.27.mlp.down_proj": 1963,
+ "model.layers.28.self_attn.q_proj": 1409,
+ "model.layers.28.self_attn.k_proj": 1410,
+ "model.layers.28.self_attn.v_proj": 1408,
+ "model.layers.28.self_attn.o_proj": 1473,
+ "model.layers.28.mlp.gate_proj": 2010,
+ "model.layers.28.mlp.up_proj": 2017,
+ "model.layers.28.mlp.down_proj": 1955,
+ "model.layers.29.self_attn.q_proj": 1397,
+ "model.layers.29.self_attn.k_proj": 1398,
+ "model.layers.29.self_attn.v_proj": 1399,
+ "model.layers.29.self_attn.o_proj": 1464,
+ "model.layers.29.mlp.gate_proj": 2002,
+ "model.layers.29.mlp.up_proj": 2009,
+ "model.layers.29.mlp.down_proj": 1948,
+ "model.layers.30.self_attn.q_proj": 1389,
+ "model.layers.30.self_attn.k_proj": 1391,
+ "model.layers.30.self_attn.v_proj": 1385,
+ "model.layers.30.self_attn.o_proj": 1436,
+ "model.layers.30.mlp.gate_proj": 1993,
+ "model.layers.30.mlp.up_proj": 2001,
+ "model.layers.30.mlp.down_proj": 1941,
+ "model.layers.31.self_attn.q_proj": 1396,
+ "model.layers.31.self_attn.k_proj": 1396,
+ "model.layers.31.self_attn.v_proj": 1395,
+ "model.layers.31.self_attn.o_proj": 1446,
+ "model.layers.31.mlp.gate_proj": 1985,
+ "model.layers.31.mlp.up_proj": 1993,
+ "model.layers.31.mlp.down_proj": 1936,
+ "model.layers.32.self_attn.q_proj": 1393,
+ "model.layers.32.self_attn.k_proj": 1395,
+ "model.layers.32.self_attn.v_proj": 1383,
+ "model.layers.32.self_attn.o_proj": 1429,
+ "model.layers.32.mlp.gate_proj": 1979,
+ "model.layers.32.mlp.up_proj": 1985,
+ "model.layers.32.mlp.down_proj": 1929,
+ "model.layers.33.self_attn.q_proj": 1399,
+ "model.layers.33.self_attn.k_proj": 1400,
+ "model.layers.33.self_attn.v_proj": 1390,
+ "model.layers.33.self_attn.o_proj": 1430,
+ "model.layers.33.mlp.gate_proj": 1973,
+ "model.layers.33.mlp.up_proj": 1978,
+ "model.layers.33.mlp.down_proj": 1921,
+ "model.layers.34.self_attn.q_proj": 1401,
+ "model.layers.34.self_attn.k_proj": 1401,
+ "model.layers.34.self_attn.v_proj": 1383,
+ "model.layers.34.self_attn.o_proj": 1424,
+ "model.layers.34.mlp.gate_proj": 1968,
+ "model.layers.34.mlp.up_proj": 1972,
+ "model.layers.34.mlp.down_proj": 1910,
+ "model.layers.35.self_attn.q_proj": 1407,
+ "model.layers.35.self_attn.k_proj": 1407,
+ "model.layers.35.self_attn.v_proj": 1388,
+ "model.layers.35.self_attn.o_proj": 1412,
+ "model.layers.35.mlp.gate_proj": 1964,
+ "model.layers.35.mlp.up_proj": 1966,
+ "model.layers.35.mlp.down_proj": 1897,
+ "model.layers.36.self_attn.q_proj": 1417,
+ "model.layers.36.self_attn.k_proj": 1416,
+ "model.layers.36.self_attn.v_proj": 1392,
+ "model.layers.36.self_attn.o_proj": 1380,
+ "model.layers.36.mlp.gate_proj": 1962,
+ "model.layers.36.mlp.up_proj": 1963,
+ "model.layers.36.mlp.down_proj": 1882,
+ "model.layers.37.self_attn.q_proj": 1438,
+ "model.layers.37.self_attn.k_proj": 1436,
+ "model.layers.37.self_attn.v_proj": 1398,
+ "model.layers.37.self_attn.o_proj": 1391,
+ "model.layers.37.mlp.gate_proj": 1964,
+ "model.layers.37.mlp.up_proj": 1966,
+ "model.layers.37.mlp.down_proj": 1877,
+ "model.layers.38.self_attn.q_proj": 1446,
+ "model.layers.38.self_attn.k_proj": 1442,
+ "model.layers.38.self_attn.v_proj": 1392,
+ "model.layers.38.self_attn.o_proj": 1376,
+ "model.layers.38.mlp.gate_proj": 1971,
+ "model.layers.38.mlp.up_proj": 1978,
+ "model.layers.38.mlp.down_proj": 1877,
+ "model.layers.39.self_attn.q_proj": 1470,
+ "model.layers.39.self_attn.k_proj": 1466,
+ "model.layers.39.self_attn.v_proj": 1423,
+ "model.layers.39.self_attn.o_proj": 1389,
+ "model.layers.39.mlp.gate_proj": 1998,
+ "model.layers.39.mlp.up_proj": 2011,
+ "model.layers.39.mlp.down_proj": 1865
+}

svd-llm-v2-13b-unhealed/state_dict_fp16.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f0dbe9c89696d52b78344dc7564c8c5bc6bb57ecf7cd95edaa26e6738db87e23
+size 15480786677

svd-llm-v2-13b/README.md

@@ -0,0 +1,29 @@
+---
+license: llama2
+base_model: NousResearch/Llama-2-13b-chat-hf
+tags: [svd, low-rank, model-compression, llama2, knowledge-distillation]
+---
+
+# SVD-LLM V2 @ 7.74B (59.47% kept of LLaMA-2-13B-chat), KD-healed 60M tokens
+
+SVD-LLM V2 (NAACL 2025): heterogeneous ratio allocation (Alg. 1) + loss-optimized truncation (Alg. 2). Official code never released; our faithful implementation (see per-layer ranks in ranks.json). Exact budget match: 7,740,286,464 vs target 7,740,282,880.
+
+- **HEALED checkpoint** (this dir). The matching **UNHEALED** pre-distillation weights are in
+  `svd-llm-v2-13b-unhealed/` — do not mix them up: same `ranks.json`, different `state_dict_fp16.pt` (see
+  `provenance.json` SHA256 here vs the unhealed dir's record).
+- Healing: pure forward-KL distillation from frozen NousResearch/Llama-2-13b-chat-hf teacher,
+  60M tokens, held-out KL **2.0268 -> 0.2858**, probe MMLU (5-shot, limit 40) **26.89% -> 44.96%**.
+- Identical recipe/protocol to the TR-3 13B and the 7B SVD baselines (see their dirs).
+
+## Load
+
+```python
+import json
+from lowrank_linear import build, materialize_all
+model = build("NousResearch/Llama-2-13b-chat-hf", json.load(open("ranks.json")),
+              "state_dict_fp16.pt", device="cuda")
+materialize_all(model)   # inference-only fusion U@V
+```
+
+Full provenance (data mixture + revisions, decontamination, SHA256 of every input and
+checkpoint, 12 x 5M milestone trajectory) in `provenance.json`.

svd-llm-v2-13b/config.json

@@ -0,0 +1,26 @@
+{
+  "_name_or_path": null,
+  "architectures": [
+    "LlamaForCausalLM"
+  ],
+  "bos_token_id": 1,
+  "eos_token_id": 2,
+  "hidden_act": "silu",
+  "hidden_size": 5120,
+  "initializer_range": 0.02,
+  "intermediate_size": 13824,
+  "max_position_embeddings": 4096,
+  "model_type": "llama",
+  "num_attention_heads": 40,
+  "num_hidden_layers": 40,
+  "num_key_value_heads": 40,
+  "pad_token_id": 0,
+  "pretraining_tp": 1,
+  "rms_norm_eps": 1e-05,
+  "rope_scaling": null,
+  "tie_word_embeddings": false,
+  "torch_dtype": "float16",
+  "transformers_version": "4.31.0.dev0",
+  "use_cache": true,
+  "vocab_size": 32000
+}

svd-llm-v2-13b/lowrank_linear.py

@@ -0,0 +1,83 @@
+"""Minimal standalone loader for the SVD-baseline compressed LLaMA-2-7B checkpoints.
+
+No dependency beyond torch + transformers. The state_dict stores every compressed
+projection as two factors of a LowRankLinear (A: out x k, B: k x in); ranks.json
+holds the per-layer integer rank of each projection. Untouched parameters
+(embeddings, norms) keep their stock transformers names.
+
+Usage:
+    import json, torch
+    from lowrank_linear import build
+    model = build(
+        base_id="NousResearch/Llama-2-7b-chat-hf",
+        ranks=json.load(open("ranks.json")),
+        state_dict_path="state_dict_fp16.pt",
+        device="cuda", dtype=torch.float16,
+    )
+    tokenizer = AutoTokenizer.from_pretrained("NousResearch/Llama-2-7b-chat-hf")
+    out = model.generate(tokenizer("2+2=", return_tensors="pt").to("cuda").input_ids,
+                         max_new_tokens=16, do_sample=False)
+    print(tokenizer.decode(out[0]))
+"""
+
+from __future__ import annotations
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+
+class LowRankLinear(nn.Module):
+    """y = A (B x).  A: out x r, B: r x in.  (LLaMA projections carry no bias.)"""
+
+    def __init__(self, A: torch.Tensor, B: torch.Tensor):
+        super().__init__()
+        self.A = nn.Parameter(A)
+        self.B = nn.Parameter(B)
+
+    def forward(self, x):
+        return F.linear(F.linear(x, self.B), self.A)
+
+    @torch.no_grad()
+    def materialize_(self) -> nn.Linear:
+        """Replace this module with the dense W = A @ B (inference only; never call
+        while training -- it cuts the gradient path)."""
+        W = (self.A @ self.B).to(self.A.dtype)
+        dense = nn.Linear(self.B.shape[1], self.A.shape[0], bias=False,
+                          device=W.device, dtype=W.dtype)
+        dense.weight.data = W
+        return dense
+
+
+def build(base_id: str, ranks: dict, state_dict_path: str,
+          device: str = "cuda", dtype=torch.float16):
+    from transformers import AutoModelForCausalLM
+
+    model = AutoModelForCausalLM.from_pretrained(base_id, torch_dtype=dtype)
+    search = {}
+    for name, mod in model.named_modules():
+        if isinstance(mod, nn.Linear):
+            search[name] = mod
+    for path, k in ranks.items():
+        out_f, in_f = search[path].weight.shape
+        model_mod = model
+        for p in path.split(".")[:-1]:
+            model_mod = getattr(model_mod, p)
+        setattr(model_mod, path.split(".")[-1],
+                LowRankLinear(torch.zeros(out_f, int(k), dtype=dtype),
+                              torch.zeros(int(k), in_f, dtype=dtype)))
+    sd = torch.load(state_dict_path, map_location="cpu", weights_only=True)
+    model.load_state_dict(sd, strict=True)
+    del sd
+    return model.to(device)
+
+
+def materialize_all(model):
+    """Fuse every LowRankLinear back into a dense nn.Linear in place."""
+    for name, mod in list(model.named_modules()):
+        if isinstance(mod, LowRankLinear):
+            parent = model
+            for p in name.split(".")[:-1]:
+                parent = getattr(parent, p)
+            setattr(parent, name.split(".")[-1], mod.materialize_())
+    return model

svd-llm-v2-13b/provenance.json

@@ -0,0 +1,199 @@
+{
+  "method": "KD healing: alpha*T^2*KL(teacher||student) + (1-alpha)*CE",
+  "student_init": "/data1/rkp00022/zixuan/projects/TR3_main/checkpoints/svdllm2_r04158_tr3fmt/state_dict_fp16.pt",
+  "teacher": "NousResearch/Llama-2-13b-chat-hf",
+  "alpha": 1.0,
+  "temperature": 1.0,
+  "lr": 2e-05,
+  "tokens_seen": 59998208,
+  "epochs": 0.228,
+  "global_batch_tokens": 32768,
+  "dataset_mixture_pre_exclusion": {
+    "alpaca": 50000,
+    "openhermes": 200000,
+    "ultrachat": 100000
+  },
+  "dataset_revisions": {
+    "tatsu-lab/alpaca": "dce01c9b08f87459cf36a430d809084718273017",
+    "teknium/OpenHermes-2.5": "b82037821055c377bed0d495e72e46de3bc72e84",
+    "HuggingFaceH4/ultrachat_200k": "8049631c405ae6576f93f445c6b8166f76f5505a"
+  },
+  "canonical_pool_excluded_texts": 10000,
+  "train_texts_after_exclusion": 356000,
+  "chat_texts": 340000,
+  "mmlu_aux_texts_kept": 16000,
+  "mmlu_aux_split": "cais/mmlu auxiliary_train (official TRAIN split, disjoint from the evaluated test split; audit verified 0/99842 exact overlap)",
+  "mmlu_aux_stats": {
+    "requested_groups": 16000,
+    "rendered_groups": 16000,
+    "shots": 5,
+    "rows_total": 99842,
+    "rows_dropped_by_decontam": 71,
+    "rows_clean": 99771,
+    "revision": "c30699e8356da336a370243923dbaf21066bb9fe",
+    "format": "MMLU-shaped (NOT byte-identical to lm-eval; auxiliary_train has no subject)"
+  },
+  "gold_label_supervision": false,
+  "attribution_note": "alpha=1.0 => pure KD, no gold-label CE, so any MMLU change is attributable to distillation on in-domain text. alpha<1 adds CE on text containing answers = labeled MCQ training, which is a DIFFERENT and weaker claim.",
+  "train_tokens_total": 263112704,
+  "decontamination": {
+    "removed": {
+      "alpaca": 11,
+      "openhermes": 120,
+      "ultrachat": 33
+    }
+  },
+  "start_kl": 2.026848241686821,
+  "final_kl": 0.28583861887454987,
+  "start_mmlu": 0.268859649122807,
+  "final_mmlu": 0.44956140350877194,
+  "history": [
+    {
+      "tokens_m": 5.01,
+      "step": 153,
+      "held_out_kl": 0.5490630865097046,
+      "mmlu": 0.42105263157894735
+    },
+    {
+      "tokens_m": 10.03,
+      "step": 306,
+      "held_out_kl": 0.5244443118572235,
+      "mmlu": 0.4337719298245614
+    },
+    {
+      "tokens_m": 15.01,
+      "step": 458,
+      "held_out_kl": 0.49302633106708527,
+      "mmlu": 0.4337719298245614
+    },
+    {
+      "tokens_m": 20.02,
+      "step": 611,
+      "held_out_kl": 0.47258753702044487,
+      "mmlu": 0.40964912280701754
+    },
+    {
+      "tokens_m": 25.0,
+      "step": 763,
+      "held_out_kl": 0.4351510964334011,
+      "mmlu": 0.4236842105263158
+    },
+    {
+      "tokens_m": 30.02,
+      "step": 916,
+      "held_out_kl": 0.39489416405558586,
+      "mmlu": 0.42412280701754385
+    },
+    {
+      "tokens_m": 35.03,
+      "step": 1069,
+      "held_out_kl": 0.36299458146095276,
+      "mmlu": 0.4359649122807018
+    },
+    {
+      "tokens_m": 40.01,
+      "step": 1221,
+      "held_out_kl": 0.33124174922704697,
+      "mmlu": 0.4219298245614035
+    },
+    {
+      "tokens_m": 45.02,
+      "step": 1374,
+      "held_out_kl": 0.30896387435495853,
+      "mmlu": 0.4355263157894737
+    },
+    {
+      "tokens_m": 50.0,
+      "step": 1526,
+      "held_out_kl": 0.2978489622473717,
+      "mmlu": 0.44605263157894737
+    },
+    {
+      "tokens_m": 55.02,
+      "step": 1679,
+      "held_out_kl": 0.28977251797914505,
+      "mmlu": 0.44605263157894737
+    },
+    {
+      "tokens_m": 60.0,
+      "step": 1831,
+      "held_out_kl": 0.28583861887454987,
+      "mmlu": 0.44956140350877194,
+      "final": true
+    }
+  ],
+  "teacher_mmlu_same_limit": 0.5442982456140351,
+  "checkpoint": "/data1/rkp00022/zixuan/projects/TR3_main/checkpoints/kdheal_svdllm2r04158_h1/state_dict_fp16.pt",
+  "sha256": "1d7cee642e764d6041c21aa1c9177635d9dcc48da03de3d0869c834348a4a676",
+  "seed": 1234,
+  "mmlu_limit_per_subtask": 40,
+  "teacher_revision": "d73f5fa9c4bc135502e04c27b39660747172d76b",
+  "student_init_sha256": "f0dbe9c89696d52b78344dc7564c8c5bc6bb57ecf7cd95edaa26e6738db87e23",
+  "ranks_file": "/data1/rkp00022/zixuan/projects/TR3_main/checkpoints/svdllm2_r04158_tr3fmt/ranks.json",
+  "ranks_sha256": "deedef0f4a8e33a06a5584b6e5417b5b8262342a1cfbcf797afe12e2fbfb9cc5",
+  "input_identity_gate": {
+    "student_ckpt": {
+      "key": "svdllm2_r04158_tr3fmt",
+      "sha256": "f0dbe9c89696d52b78344dc7564c8c5bc6bb57ecf7cd95edaa26e6738db87e23",
+      "status": "unregistered_allowed"
+    },
+    "ranks": {
+      "key": "svdllm2_r04158_tr3fmt",
+      "sha256": "deedef0f4a8e33a06a5584b6e5417b5b8262342a1cfbcf797afe12e2fbfb9cc5",
+      "status": "unregistered_allowed"
+    }
+  },
+  "train_seed": 1234,
+  "mmlu_seed": 1234,
+  "dynamic_batch": false,
+  "dbl": null,
+  "snapshot_milestones": true,
+  "snapshots": [
+    "(PosixPath('/data1/rkp00022/zixuan/projects/TR3_main/checkpoints/kdheal_svdllm2r04158_h1_snap15m'), 15.007744)",
+    "(PosixPath('/data1/rkp00022/zixuan/projects/TR3_main/checkpoints/kdheal_svdllm2r04158_h1_snap30m'), 30.015488)",
+    "(PosixPath('/data1/rkp00022/zixuan/projects/TR3_main/checkpoints/kdheal_svdllm2r04158_h1_snap45m'), 45.023232)",
+    "(PosixPath('/data1/rkp00022/zixuan/projects/TR3_main/checkpoints/kdheal_svdllm2r04158_h1_snap50m'), 50.003968)",
+    "(PosixPath('/data1/rkp00022/zixuan/projects/TR3_main/checkpoints/kdheal_svdllm2r04158_h1_snap55m'), 55.017472)"
+  ],
+  "stream_skip_tokens_m": 0.0,
+  "stream_blocks_consumed": [
+    0,
+    58592
+  ],
+  "stream_note": "the block order is a seeded permutation consumed from index 0; stages sharing a --seed with stream_skip=0 re-train on the SAME blocks. [start, end] block indices are recorded so any continuation can be checked for overlap with what earlier stages consumed.",
+  "student_master_dtype": "bf16",
+  "fake_quant": null,
+  "rebalance_report": null,
+  "student_param_count": 7740286464,
+  "micro": 8,
+  "grad_accum": 4,
+  "seq_len": 1024,
+  "held_out_source": "canonical: mixture 2000/4000/4000, seed 1234, seq 1024, blocks 256:264 (identical to the compression gates)",
+  "decontam_stats": {
+    "mmlu": {
+      "items": 14042,
+      "new_ngrams": 764261,
+      "revision": "c30699e8356da336a370243923dbaf21066bb9fe"
+    },
+    "hellaswag": {
+      "items": 10042,
+      "new_ngrams": 1334302,
+      "revision": "218ec52e09a7e7462a5400043bb9a69a41d06b76"
+    },
+    "boolq": {
+      "items": 3270,
+      "new_ngrams": 276806,
+      "revision": "3de24cf8022e94f4ee4b9d55a6f539891524d646"
+    },
+    "triviaqa": {
+      "items": 17944,
+      "new_ngrams": 32608,
+      "revision": "0f7faf33a3908546c6fd5b73a660e0f8ff173c2f"
+    },
+    "gsm8k": {
+      "items": 1319,
+      "new_ngrams": 46281,
+      "revision": "740312add88f781978c0658806c59bc2815b9866"
+    }
+  }
+}

svd-llm-v2-13b/ranks.json

@@ -0,0 +1,282 @@
+{
+ "model.layers.0.self_attn.q_proj": 2560,
+ "model.layers.0.self_attn.k_proj": 2560,
+ "model.layers.0.self_attn.v_proj": 2560,
+ "model.layers.0.self_attn.o_proj": 2560,
+ "model.layers.0.mlp.gate_proj": 3237,
+ "model.layers.0.mlp.up_proj": 3231,
+ "model.layers.0.mlp.down_proj": 3674,
+ "model.layers.1.self_attn.q_proj": 2514,
+ "model.layers.1.self_attn.k_proj": 2553,
+ "model.layers.1.self_attn.v_proj": 2288,
+ "model.layers.1.self_attn.o_proj": 2560,
+ "model.layers.1.mlp.gate_proj": 2862,
+ "model.layers.1.mlp.up_proj": 2816,
+ "model.layers.1.mlp.down_proj": 2909,
+ "model.layers.2.self_attn.q_proj": 1842,
+ "model.layers.2.self_attn.k_proj": 1847,
+ "model.layers.2.self_attn.v_proj": 1839,
+ "model.layers.2.self_attn.o_proj": 2126,
+ "model.layers.2.mlp.gate_proj": 2506,
+ "model.layers.2.mlp.up_proj": 2500,
+ "model.layers.2.mlp.down_proj": 2665,
+ "model.layers.3.self_attn.q_proj": 1735,
+ "model.layers.3.self_attn.k_proj": 1735,
+ "model.layers.3.self_attn.v_proj": 1742,
+ "model.layers.3.self_attn.o_proj": 2020,
+ "model.layers.3.mlp.gate_proj": 2363,
+ "model.layers.3.mlp.up_proj": 2369,
+ "model.layers.3.mlp.down_proj": 2569,
+ "model.layers.4.self_attn.q_proj": 1618,
+ "model.layers.4.self_attn.k_proj": 1618,
+ "model.layers.4.self_attn.v_proj": 1651,
+ "model.layers.4.self_attn.o_proj": 1888,
+ "model.layers.4.mlp.gate_proj": 2302,
+ "model.layers.4.mlp.up_proj": 2310,
+ "model.layers.4.mlp.down_proj": 2471,
+ "model.layers.5.self_attn.q_proj": 1603,
+ "model.layers.5.self_attn.k_proj": 1601,
+ "model.layers.5.self_attn.v_proj": 1624,
+ "model.layers.5.self_attn.o_proj": 1818,
+ "model.layers.5.mlp.gate_proj": 2253,
+ "model.layers.5.mlp.up_proj": 2262,
+ "model.layers.5.mlp.down_proj": 2388,
+ "model.layers.6.self_attn.q_proj": 1555,
+ "model.layers.6.self_attn.k_proj": 1551,
+ "model.layers.6.self_attn.v_proj": 1588,
+ "model.layers.6.self_attn.o_proj": 1778,
+ "model.layers.6.mlp.gate_proj": 2218,
+ "model.layers.6.mlp.up_proj": 2227,
+ "model.layers.6.mlp.down_proj": 2327,
+ "model.layers.7.self_attn.q_proj": 1556,
+ "model.layers.7.self_attn.k_proj": 1552,
+ "model.layers.7.self_attn.v_proj": 1590,
+ "model.layers.7.self_attn.o_proj": 1701,
+ "model.layers.7.mlp.gate_proj": 2199,
+ "model.layers.7.mlp.up_proj": 2208,
+ "model.layers.7.mlp.down_proj": 2277,
+ "model.layers.8.self_attn.q_proj": 1535,
+ "model.layers.8.self_attn.k_proj": 1531,
+ "model.layers.8.self_attn.v_proj": 1565,
+ "model.layers.8.self_attn.o_proj": 1663,
+ "model.layers.8.mlp.gate_proj": 2184,
+ "model.layers.8.mlp.up_proj": 2190,
+ "model.layers.8.mlp.down_proj": 2244,
+ "model.layers.9.self_attn.q_proj": 1504,
+ "model.layers.9.self_attn.k_proj": 1501,
+ "model.layers.9.self_attn.v_proj": 1544,
+ "model.layers.9.self_attn.o_proj": 1630,
+ "model.layers.9.mlp.gate_proj": 2169,
+ "model.layers.9.mlp.up_proj": 2170,
+ "model.layers.9.mlp.down_proj": 2219,
+ "model.layers.10.self_attn.q_proj": 1507,
+ "model.layers.10.self_attn.k_proj": 1506,
+ "model.layers.10.self_attn.v_proj": 1551,
+ "model.layers.10.self_attn.o_proj": 1583,
+ "model.layers.10.mlp.gate_proj": 2166,
+ "model.layers.10.mlp.up_proj": 2164,
+ "model.layers.10.mlp.down_proj": 2197,
+ "model.layers.11.self_attn.q_proj": 1509,
+ "model.layers.11.self_attn.k_proj": 1503,
+ "model.layers.11.self_attn.v_proj": 1554,
+ "model.layers.11.self_attn.o_proj": 1577,
+ "model.layers.11.mlp.gate_proj": 2165,
+ "model.layers.11.mlp.up_proj": 2159,
+ "model.layers.11.mlp.down_proj": 2180,
+ "model.layers.12.self_attn.q_proj": 1501,
+ "model.layers.12.self_attn.k_proj": 1495,
+ "model.layers.12.self_attn.v_proj": 1534,
+ "model.layers.12.self_attn.o_proj": 1552,
+ "model.layers.12.mlp.gate_proj": 2158,
+ "model.layers.12.mlp.up_proj": 2148,
+ "model.layers.12.mlp.down_proj": 2164,
+ "model.layers.13.self_attn.q_proj": 1494,
+ "model.layers.13.self_attn.k_proj": 1494,
+ "model.layers.13.self_attn.v_proj": 1514,
+ "model.layers.13.self_attn.o_proj": 1534,
+ "model.layers.13.mlp.gate_proj": 2149,
+ "model.layers.13.mlp.up_proj": 2138,
+ "model.layers.13.mlp.down_proj": 2142,
+ "model.layers.14.self_attn.q_proj": 1482,
+ "model.layers.14.self_attn.k_proj": 1480,
+ "model.layers.14.self_attn.v_proj": 1512,
+ "model.layers.14.self_attn.o_proj": 1499,
+ "model.layers.14.mlp.gate_proj": 2141,
+ "model.layers.14.mlp.up_proj": 2129,
+ "model.layers.14.mlp.down_proj": 2122,
+ "model.layers.15.self_attn.q_proj": 1486,
+ "model.layers.15.self_attn.k_proj": 1482,
+ "model.layers.15.self_attn.v_proj": 1507,
+ "model.layers.15.self_attn.o_proj": 1512,
+ "model.layers.15.mlp.gate_proj": 2131,
+ "model.layers.15.mlp.up_proj": 2117,
+ "model.layers.15.mlp.down_proj": 2103,
+ "model.layers.16.self_attn.q_proj": 1476,
+ "model.layers.16.self_attn.k_proj": 1471,
+ "model.layers.16.self_attn.v_proj": 1490,
+ "model.layers.16.self_attn.o_proj": 1504,
+ "model.layers.16.mlp.gate_proj": 2121,
+ "model.layers.16.mlp.up_proj": 2107,
+ "model.layers.16.mlp.down_proj": 2077,
+ "model.layers.17.self_attn.q_proj": 1468,
+ "model.layers.17.self_attn.k_proj": 1466,
+ "model.layers.17.self_attn.v_proj": 1482,
+ "model.layers.17.self_attn.o_proj": 1491,
+ "model.layers.17.mlp.gate_proj": 2108,
+ "model.layers.17.mlp.up_proj": 2098,
+ "model.layers.17.mlp.down_proj": 2066,
+ "model.layers.18.self_attn.q_proj": 1455,
+ "model.layers.18.self_attn.k_proj": 1453,
+ "model.layers.18.self_attn.v_proj": 1469,
+ "model.layers.18.self_attn.o_proj": 1498,
+ "model.layers.18.mlp.gate_proj": 2094,
+ "model.layers.18.mlp.up_proj": 2088,
+ "model.layers.18.mlp.down_proj": 2058,
+ "model.layers.19.self_attn.q_proj": 1461,
+ "model.layers.19.self_attn.k_proj": 1460,
+ "model.layers.19.self_attn.v_proj": 1468,
+ "model.layers.19.self_attn.o_proj": 1499,
+ "model.layers.19.mlp.gate_proj": 2083,
+ "model.layers.19.mlp.up_proj": 2079,
+ "model.layers.19.mlp.down_proj": 2044,
+ "model.layers.20.self_attn.q_proj": 1462,
+ "model.layers.20.self_attn.k_proj": 1462,
+ "model.layers.20.self_attn.v_proj": 1474,
+ "model.layers.20.self_attn.o_proj": 1480,
+ "model.layers.20.mlp.gate_proj": 2077,
+ "model.layers.20.mlp.up_proj": 2074,
+ "model.layers.20.mlp.down_proj": 2030,
+ "model.layers.21.self_attn.q_proj": 1464,
+ "model.layers.21.self_attn.k_proj": 1464,
+ "model.layers.21.self_attn.v_proj": 1463,
+ "model.layers.21.self_attn.o_proj": 1460,
+ "model.layers.21.mlp.gate_proj": 2072,
+ "model.layers.21.mlp.up_proj": 2072,
+ "model.layers.21.mlp.down_proj": 2010,
+ "model.layers.22.self_attn.q_proj": 1436,
+ "model.layers.22.self_attn.k_proj": 1437,
+ "model.layers.22.self_attn.v_proj": 1442,
+ "model.layers.22.self_attn.o_proj": 1491,
+ "model.layers.22.mlp.gate_proj": 2061,
+ "model.layers.22.mlp.up_proj": 2063,
+ "model.layers.22.mlp.down_proj": 2003,
+ "model.layers.23.self_attn.q_proj": 1442,
+ "model.layers.23.self_attn.k_proj": 1443,
+ "model.layers.23.self_attn.v_proj": 1442,
+ "model.layers.23.self_attn.o_proj": 1488,
+ "model.layers.23.mlp.gate_proj": 2051,
+ "model.layers.23.mlp.up_proj": 2056,
+ "model.layers.23.mlp.down_proj": 1994,
+ "model.layers.24.self_attn.q_proj": 1433,
+ "model.layers.24.self_attn.k_proj": 1435,
+ "model.layers.24.self_attn.v_proj": 1436,
+ "model.layers.24.self_attn.o_proj": 1495,
+ "model.layers.24.mlp.gate_proj": 2044,
+ "model.layers.24.mlp.up_proj": 2050,
+ "model.layers.24.mlp.down_proj": 1988,
+ "model.layers.25.self_attn.q_proj": 1428,
+ "model.layers.25.self_attn.k_proj": 1428,
+ "model.layers.25.self_attn.v_proj": 1425,
+ "model.layers.25.self_attn.o_proj": 1475,
+ "model.layers.25.mlp.gate_proj": 2035,
+ "model.layers.25.mlp.up_proj": 2042,
+ "model.layers.25.mlp.down_proj": 1978,
+ "model.layers.26.self_attn.q_proj": 1424,
+ "model.layers.26.self_attn.k_proj": 1424,
+ "model.layers.26.self_attn.v_proj": 1421,
+ "model.layers.26.self_attn.o_proj": 1474,
+ "model.layers.26.mlp.gate_proj": 2027,
+ "model.layers.26.mlp.up_proj": 2035,
+ "model.layers.26.mlp.down_proj": 1969,
+ "model.layers.27.self_attn.q_proj": 1408,
+ "model.layers.27.self_attn.k_proj": 1409,
+ "model.layers.27.self_attn.v_proj": 1410,
+ "model.layers.27.self_attn.o_proj": 1474,
+ "model.layers.27.mlp.gate_proj": 2019,
+ "model.layers.27.mlp.up_proj": 2027,
+ "model.layers.27.mlp.down_proj": 1963,
+ "model.layers.28.self_attn.q_proj": 1409,
+ "model.layers.28.self_attn.k_proj": 1410,
+ "model.layers.28.self_attn.v_proj": 1408,
+ "model.layers.28.self_attn.o_proj": 1473,
+ "model.layers.28.mlp.gate_proj": 2010,
+ "model.layers.28.mlp.up_proj": 2017,
+ "model.layers.28.mlp.down_proj": 1955,
+ "model.layers.29.self_attn.q_proj": 1397,
+ "model.layers.29.self_attn.k_proj": 1398,
+ "model.layers.29.self_attn.v_proj": 1399,
+ "model.layers.29.self_attn.o_proj": 1464,
+ "model.layers.29.mlp.gate_proj": 2002,
+ "model.layers.29.mlp.up_proj": 2009,
+ "model.layers.29.mlp.down_proj": 1948,
+ "model.layers.30.self_attn.q_proj": 1389,
+ "model.layers.30.self_attn.k_proj": 1391,
+ "model.layers.30.self_attn.v_proj": 1385,
+ "model.layers.30.self_attn.o_proj": 1436,
+ "model.layers.30.mlp.gate_proj": 1993,
+ "model.layers.30.mlp.up_proj": 2001,
+ "model.layers.30.mlp.down_proj": 1941,
+ "model.layers.31.self_attn.q_proj": 1396,
+ "model.layers.31.self_attn.k_proj": 1396,
+ "model.layers.31.self_attn.v_proj": 1395,
+ "model.layers.31.self_attn.o_proj": 1446,
+ "model.layers.31.mlp.gate_proj": 1985,
+ "model.layers.31.mlp.up_proj": 1993,
+ "model.layers.31.mlp.down_proj": 1936,
+ "model.layers.32.self_attn.q_proj": 1393,
+ "model.layers.32.self_attn.k_proj": 1395,
+ "model.layers.32.self_attn.v_proj": 1383,
+ "model.layers.32.self_attn.o_proj": 1429,
+ "model.layers.32.mlp.gate_proj": 1979,
+ "model.layers.32.mlp.up_proj": 1985,
+ "model.layers.32.mlp.down_proj": 1929,
+ "model.layers.33.self_attn.q_proj": 1399,
+ "model.layers.33.self_attn.k_proj": 1400,
+ "model.layers.33.self_attn.v_proj": 1390,
+ "model.layers.33.self_attn.o_proj": 1430,
+ "model.layers.33.mlp.gate_proj": 1973,
+ "model.layers.33.mlp.up_proj": 1978,
+ "model.layers.33.mlp.down_proj": 1921,
+ "model.layers.34.self_attn.q_proj": 1401,
+ "model.layers.34.self_attn.k_proj": 1401,
+ "model.layers.34.self_attn.v_proj": 1383,
+ "model.layers.34.self_attn.o_proj": 1424,
+ "model.layers.34.mlp.gate_proj": 1968,
+ "model.layers.34.mlp.up_proj": 1972,
+ "model.layers.34.mlp.down_proj": 1910,
+ "model.layers.35.self_attn.q_proj": 1407,
+ "model.layers.35.self_attn.k_proj": 1407,
+ "model.layers.35.self_attn.v_proj": 1388,
+ "model.layers.35.self_attn.o_proj": 1412,
+ "model.layers.35.mlp.gate_proj": 1964,
+ "model.layers.35.mlp.up_proj": 1966,
+ "model.layers.35.mlp.down_proj": 1897,
+ "model.layers.36.self_attn.q_proj": 1417,
+ "model.layers.36.self_attn.k_proj": 1416,
+ "model.layers.36.self_attn.v_proj": 1392,
+ "model.layers.36.self_attn.o_proj": 1380,
+ "model.layers.36.mlp.gate_proj": 1962,
+ "model.layers.36.mlp.up_proj": 1963,
+ "model.layers.36.mlp.down_proj": 1882,
+ "model.layers.37.self_attn.q_proj": 1438,
+ "model.layers.37.self_attn.k_proj": 1436,
+ "model.layers.37.self_attn.v_proj": 1398,
+ "model.layers.37.self_attn.o_proj": 1391,
+ "model.layers.37.mlp.gate_proj": 1964,
+ "model.layers.37.mlp.up_proj": 1966,
+ "model.layers.37.mlp.down_proj": 1877,
+ "model.layers.38.self_attn.q_proj": 1446,
+ "model.layers.38.self_attn.k_proj": 1442,
+ "model.layers.38.self_attn.v_proj": 1392,
+ "model.layers.38.self_attn.o_proj": 1376,
+ "model.layers.38.mlp.gate_proj": 1971,
+ "model.layers.38.mlp.up_proj": 1978,
+ "model.layers.38.mlp.down_proj": 1877,
+ "model.layers.39.self_attn.q_proj": 1470,
+ "model.layers.39.self_attn.k_proj": 1466,
+ "model.layers.39.self_attn.v_proj": 1423,
+ "model.layers.39.self_attn.o_proj": 1389,
+ "model.layers.39.mlp.gate_proj": 1998,
+ "model.layers.39.mlp.up_proj": 2011,
+ "model.layers.39.mlp.down_proj": 1865
+}

svd-llm-v2-13b/state_dict_fp16.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1d7cee642e764d6041c21aa1c9177635d9dcc48da03de3d0869c834348a4a676
+size 15480788552