ManniX-ITA commited on
Commit
5f84815
·
verified ·
1 Parent(s): 9ec0ce0

Add files using upload-large-folder tool

Browse files
.caliber/learning/current-session.jsonl ADDED
@@ -0,0 +1,7 @@
 
 
 
 
 
 
 
 
1
+ {"timestamp":"2026-05-04T06:49:50.467Z","session_id":"4208dc59-08bc-4b1e-8d4e-8ec3407fd4ac","hook_event_name":"PostToolUse","tool_name":"Bash","tool_input":{"command":"kill 52848 2>/dev/null; sleep 2; pgrep -af \"hf upload\" | grep -v claude | head; rm -rf /srv/dev-disk-by-uuid-f8b1803e-334f-4f4b-af3b-f802bb6883c5/backup_models/4b_phase1/merged/v2i-jv-base-task-arith/.caliber /srv/dev-disk-by-uuid-f8b1803e-334f-4f4b-af3b-f802bb6883c5/backup_models/4b_phase1/merged/v2i-jv-base-task-arith/.merge_done; ls -la /srv/dev-disk-by-uuid-f8b1803e-334f-4f4b-af3b-f802bb6883c5/backup_models/4b_phase1/merged/v2i-jv-base-task-arith/ | head","description":"Kill upload, clean dir"},"tool_response":{"stdout":"total 9121028\ndrwxr-sr-x 3 root users 323 May 4 08:49 .\ndrwxr-sr-x 39 root users 4096 May 4 01:44 ..\ndrwxr-sr-x 3 root users 33 May 4 07:23 .cache\n-rw-r--r-- 1 root users 4047 May 1 14:24 chat_template.jinja\n-rw-r--r-- 1 root users 3471 May 1 14:24 config.json\n-rw-r--r-- 1 root users 5376584768 May 3 18:45 model-00001-of-00002.safetensors\n-rw-r--r-- 1 root users 3943243384 May 3 18:49 model-00002-of-00002.safetensors\n-rw-r--r-- 1 root users 67340 May 3 18:50 model.safetensors.index.json\n-rw-r--r-- 1 root users 1300 May 1 14:35 processor_config.json","stderr":"","interrupted":false,"isImage":false,"noOutputExpected":false},"tool_use_id":"toolu_018GzyrD2A6eTH8SFXNAsXe4","cwd":"/srv/dev-disk-by-uuid-f8b1803e-334f-4f4b-af3b-f802bb6883c5/backup_models/4b_phase1/merged/v2i-jv-base-task-arith"}
2
+ {"timestamp":"2026-05-04T06:50:01.952Z","session_id":"4208dc59-08bc-4b1e-8d4e-8ec3407fd4ac","hook_event_name":"PostToolUse","tool_name":"Bash","tool_input":{"command":"cd /srv/dev-disk-by-uuid-f8b1803e-334f-4f4b-af3b-f802bb6883c5/backup_models/4b_phase1/merged/v2i-jv-base-task-arith; nohup hf upload-large-folder --repo-type=model --no-bars --no-report ManniX-ITA/Qwen3.5-4B-MicroCoder . > /srv/dev-disk-by-uuid-f8b1803e-334f-4f4b-af3b-f802bb6883c5/backup_models/logs/microcoder_weights_upload.log 2>&1 & echo \"PID=$!\"; disown; sleep 3; pgrep -af \"hf upload-large-folder\" | grep -v claude | head -2","description":"Restart upload, clean output"},"tool_response":{"stdout":"PID=57784\n57784 /usr/bin/python3 /usr/local/bin/hf upload-large-folder --repo-type=model --no-bars --no-report ManniX-ITA/Qwen3.5-4B-MicroCoder .","stderr":"","interrupted":false,"isImage":false,"noOutputExpected":false},"tool_use_id":"toolu_016duUfqqtzKAvy7rpBHGJt7","cwd":"/srv/dev-disk-by-uuid-f8b1803e-334f-4f4b-af3b-f802bb6883c5/backup_models/4b_phase1/merged/v2i-jv-base-task-arith"}
3
+ {"timestamp":"2026-05-04T07:02:41.683Z","session_id":"4208dc59-08bc-4b1e-8d4e-8ec3407fd4ac","hook_event_name":"UserPromptSubmit","prompt_content":"Diving in now in which sense? why you stopped then? are you waiting for something?","cwd":"/srv/dev-disk-by-uuid-f8b1803e-334f-4f4b-af3b-f802bb6883c5/backup_models/4b_phase1/merged/v2i-jv-base-task-arith"}
4
+ {"timestamp":"2026-05-04T07:02:47.668Z","session_id":"4208dc59-08bc-4b1e-8d4e-8ec3407fd4ac","hook_event_name":"PostToolUse","tool_name":"Bash","tool_input":{"command":"mkdir -p /shared/dev/omnimergekit/recipes/microcoder_4b /srv/dev-disk-by-uuid-f8b1803e-334f-4f4b-af3b-f802bb6883c5/backup_models/4b_phase1/pruned","description":"Create output dirs"},"tool_response":{"stdout":"","stderr":"","interrupted":false,"isImage":false,"noOutputExpected":true},"tool_use_id":"toolu_011c2n8bsYtgkKBPuG5qk834","cwd":"/srv/dev-disk-by-uuid-f8b1803e-334f-4f4b-af3b-f802bb6883c5/backup_models/4b_phase1/merged/v2i-jv-base-task-arith"}
5
+ {"timestamp":"2026-05-04T07:10:04.049Z","session_id":"4208dc59-08bc-4b1e-8d4e-8ec3407fd4ac","hook_event_name":"PostToolUse","tool_name":"Write","tool_input":{"file_path":"/shared/dev/omnimergekit/recipes/microcoder_4b/prune_local_heal.py","content":"#!/usr/bin/env python3\n\"\"\"\nprune_local_heal.py — Qwen3.5 hybrid-arch attention head pruning with\n backprop-free local healing (W_O lstsq refit).\n\nTargets the Qwen3.5 hybrid architecture which mixes:\n - full_attention layers (Qwen3_5Attention with gated query, GQA 16:4, head_dim=256)\n - linear_attention layers (Qwen3_5GatedDeltaNet, 32 v_heads / 16 k_heads, head_dim=128)\n\nTwo prune levers:\n --prune-full-frac : fraction of Q heads to drop in full_attention layers\n --prune-linear-frac : fraction of V heads to drop in linear_attention layers\n\nPhases:\n 0. Capture per-layer o_proj-input (full_attn) and out_proj-input (linear_attn)\n on a calibration set — these are the targets for healing.\n 1. Score per-head importance via Michel-style gradient on per-head scale α=1.\n 2. For each layer, in forward order, on the *current* (already-modified) model:\n a. Re-capture pre-proj input\n b. Zero rows for dropped heads in q/k/v / in_proj_qkv (and corresponding\n rows in z/b/a/conv1d/A_log/dt_bias for linear_attn)\n c. lstsq refit the kept columns of o_proj / out_proj so the post-proj\n output matches the original target\n\nThe model stays full-size on disk (mask-style); a separate physical-resize\nstep converts to a smaller model after eval validation. Mask-pruning is what\nlstsq refit operates on, and it preserves output quality identically — only\ninference compute differs.\n\"\"\"\nfrom __future__ import annotations\n\nimport argparse\nimport json\nimport os\nimport time\nfrom pathlib import Path\n\nimport torch\nimport torch.nn as nn\nimport torch.nn.functional as F\nfrom transformers import AutoModelForCausalLM, AutoTokenizer\n\n\ndef log(msg: str):\n print(f\"[{time.strftime('%H:%M:%S')}] {msg}\", flush=True)\n\n\n# ---------- calibration ----------\n\nCALIB_TEXTS = [\n # Code-leaning (HumanEval/MBPP-style prompts) + a few prose to keep general-purpose signal\n \"from typing import List\\n\\n\\ndef has_close_elements(numbers: List[float], threshold: float) -> bool:\\n \\\"\\\"\\\"Check if there are any two numbers in the list that are closer than threshold.\\\"\\\"\\\"\\n for i, a in enumerate(numbers):\\n for j, b in enumerate(numbers):\\n if i != j and abs(a - b) < threshold:\\n return True\\n return False\\n\",\n \"def fibonacci(n: int) -> int:\\n \\\"\\\"\\\"Return the n-th Fibonacci number.\\\"\\\"\\\"\\n if n < 2:\\n return n\\n a, b = 0, 1\\n for _ in range(n - 1):\\n a, b = b, a + b\\n return b\\n\",\n \"import re\\n\\n\\ndef is_palindrome(s: str) -> bool:\\n s = re.sub(r'[^A-Za-z0-9]', '', s).lower()\\n return s == s[::-1]\\n\",\n \"from collections import Counter\\n\\n\\ndef most_common_word(text: str) -> str:\\n words = text.lower().split()\\n return Counter(words).most_common(1)[0][0]\\n\",\n \"def quicksort(arr):\\n if len(arr) <= 1: return arr\\n pivot = arr[len(arr)//2]\\n return quicksort([x for x in arr if x < pivot]) + [x for x in arr if x == pivot] + quicksort([x for x in arr if x > pivot])\\n\",\n \"def merge_sorted(a, b):\\n \\\"\\\"\\\"Merge two pre-sorted lists into a sorted list.\\\"\\\"\\\"\\n out, i, j = [], 0, 0\\n while i < len(a) and j < len(b):\\n if a[i] <= b[j]: out.append(a[i]); i += 1\\n else: out.append(b[j]); j += 1\\n out.extend(a[i:]); out.extend(b[j:])\\n return out\\n\",\n \"import heapq\\n\\n\\ndef k_largest(nums, k):\\n return heapq.nlargest(k, nums)\\n\",\n \"class Stack:\\n def __init__(self): self._items = []\\n def push(self, x): self._items.append(x)\\n def pop(self): return self._items.pop() if self._items else None\\n def peek(self): return self._items[-1] if self._items else None\\n def __len__(self): return len(self._items)\\n\",\n \"def binary_search(arr, target):\\n lo, hi = 0, len(arr) - 1\\n while lo <= hi:\\n mid = (lo + hi) // 2\\n if arr[mid] == target: return mid\\n if arr[mid] < target: lo = mid + 1\\n else: hi = mid - 1\\n return -1\\n\",\n \"import json, sys\\n\\n\\ndef sum_field(path: str, field: str) -> float:\\n with open(path) as f:\\n rows = json.load(f)\\n return sum(r.get(field, 0) for r in rows)\\n\",\n \"The mitochondrion is a double-membrane-bound organelle found in most eukaryotic cells. They generate most of the cell's supply of adenosine triphosphate (ATP), used as a source of chemical energy.\",\n \"In linear algebra, the determinant is a scalar value that is a function of the entries of a square matrix. It allows characterizing some properties of the matrix and the linear map represented by it.\",\n]\n\n\ndef build_calib(tokenizer, target_tokens: int, device: str) -> torch.Tensor:\n \"\"\"Concat CALIB_TEXTS, tokenize, take first target_tokens. Returns (1, T) on device.\"\"\"\n text = \"\\n\\n\".join(CALIB_TEXTS)\n while True:\n ids = tokenizer(text, return_tensors=\"pt\").input_ids\n if ids.shape[1] >= target_tokens:\n break\n text = text + \"\\n\\n\" + text # double until enough\n ids = ids[:, :target_tokens].to(device)\n log(f\"calib: {ids.shape[1]} tokens (vocab id range {ids.min().item()}..{ids.max().item()})\")\n return ids\n\n\n# ---------- model layout helpers ----------\n\ndef get_layers(model):\n \"\"\"Return decoder layers list regardless of nesting.\"\"\"\n if hasattr(model, \"model\") and hasattr(model.model, \"layers\"):\n return model.model.layers\n if hasattr(model, \"model\") and hasattr(model.model, \"language_model\"):\n return model.model.language_model.layers\n if hasattr(model, \"language_model\"):\n return model.language_model.layers\n raise RuntimeError(\"could not locate decoder layers\")\n\n\ndef text_config(model):\n cfg = model.config\n if hasattr(cfg, \"text_config\"):\n return cfg.text_config\n return cfg\n\n\ndef layer_kind(model, idx: int) -> str:\n \"\"\"'full_attention' or 'linear_attention'.\"\"\"\n return text_config(model).layer_types[idx]\n\n\ndef is_full(layer) -> bool:\n return hasattr(layer, \"self_attn\")\n\n\ndef is_linear(layer) -> bool:\n return hasattr(layer, \"linear_attn\")\n\n\n# ---------- phase 0: capture targets ----------\n\ndef phase0_capture(model, calib_ids):\n \"\"\"\n Forward the un-modified model and capture, per attention-bearing layer:\n - full_attn: input to o_proj (the gated, concatenated head output, shape [B, T, n_q*head_dim])\n - linear_attn: input to out_proj (shape [B, T, value_dim])\n These are the targets the lstsq refit will try to reproduce post-prune.\n \"\"\"\n layers = get_layers(model)\n targets = {} # layer_idx -> tensor (cpu)\n handles = []\n\n for L, layer in enumerate(layers):\n if is_full(layer):\n proj = layer.self_attn.o_proj\n elif is_linear(layer):\n proj = layer.linear_attn.out_proj\n else:\n continue\n\n def make_hook(idx):\n def hook(module, inputs, output):\n # input is the tensor fed into o_proj/out_proj; capture and detach\n x = inputs[0].detach().to(torch.float32).cpu()\n targets[idx] = x\n return hook\n handles.append(proj.register_forward_pre_hook(lambda m, inp, idx=L: targets.__setitem__(idx, inp[0].detach().to(torch.float32).cpu())))\n\n log(f\"phase0: registered {len(handles)} pre-hooks; running forward...\")\n with torch.no_grad():\n model(calib_ids)\n\n for h in handles:\n h.remove()\n log(f\"phase0: captured {len(targets)} layer targets\")\n return targets\n\n\n# ---------- phase 1: per-head importance via Michel-style gradient ----------\n\nclass FullAttnHeadGate(nn.Module):\n \"\"\"\n Wraps Qwen3_5Attention to insert a per-Q-head scaling factor α.\n α multiplies the gated, concatenated head output before o_proj. After\n fwd+bwd, |grad α| at α=1 ranks Q-head importance.\n \"\"\"\n def __init__(self, attn, num_q_heads: int, head_dim: int):\n super().__init__()\n self.attn = attn\n self.alpha = nn.Parameter(torch.ones(num_q_heads))\n self.num_q_heads = num_q_heads\n self.head_dim = head_dim\n self._orig_o_proj_forward = attn.o_proj.forward\n\n def patched_o_proj(x: torch.Tensor) -> torch.Tensor:\n # x: [B, T, n_q*head_dim]. Reshape to [B, T, n_q, head_dim], scale, flatten.\n B, T, _ = x.shape\n x_h = x.view(B, T, self.num_q_heads, self.head_dim)\n x_h = x_h * self.alpha.view(1, 1, -1, 1)\n x = x_h.view(B, T, -1)\n return self._orig_o_proj_forward(x)\n\n attn.o_proj.forward = patched_o_proj\n\n def restore(self):\n self.attn.o_proj.forward = self._orig_o_proj_forward\n\n\nclass LinearAttnHeadGate(nn.Module):\n \"\"\"\n Wraps Qwen3_5GatedDeltaNet to insert a per-V-head scaling factor α.\n α scales the value-dim signal feeding out_proj (post-norm, pre-projection).\n \"\"\"\n def __init__(self, lin_attn, num_v_heads: int, head_v_dim: int):\n super().__init__()\n self.lin_attn = lin_attn\n self.alpha = nn.Parameter(torch.ones(num_v_heads))\n self.num_v_heads = num_v_heads\n self.head_v_dim = head_v_dim\n self._orig_out_proj_forward = lin_attn.out_proj.forward\n\n def patched_out_proj(x: torch.Tensor) -> torch.Tensor:\n # x: [B, T, value_dim] = [B, T, n_v*head_v_dim]\n B, T, _ = x.shape\n x_h = x.view(B, T, self.num_v_heads, self.head_v_dim)\n x_h = x_h * self.alpha.view(1, 1, -1, 1)\n x = x_h.view(B, T, -1)\n return self._orig_out_proj_forward(x)\n\n lin_attn.out_proj.forward = patched_out_proj\n\n def restore(self):\n self.lin_attn.out_proj.forward = self._orig_out_proj_forward\n\n\ndef phase1_importance(model, calib_ids):\n \"\"\"\n Insert α per attention-bearing layer, fwd+bwd a single CE loss (next-token),\n return |grad α| per head, per layer.\n \"\"\"\n cfg = text_config(model)\n layers = get_layers(model)\n n_q_full = cfg.num_attention_heads\n head_dim_full = cfg.head_dim\n n_v_lin = cfg.linear_num_value_heads\n head_v_dim_lin = cfg.linear_value_head_dim\n\n gates = {} # layer_idx -> gate\n for L, layer in enumerate(layers):\n if is_full(layer):\n g = FullAttnHeadGate(layer.self_attn, n_q_full, head_dim_full)\n g.alpha.requires_grad_(True)\n gates[L] = g\n elif is_linear(layer):\n g = LinearAttnHeadGate(layer.linear_attn, n_v_lin, head_v_dim_lin)\n g.alpha.requires_grad_(True)\n gates[L] = g\n\n # All non-α params: no grad\n for p in model.parameters():\n p.requires_grad_(False)\n for g in gates.values():\n g.alpha.requires_grad_(True)\n\n log(f\"phase1: gates installed on {len(gates)} layers; computing CE+bwd...\")\n model.train() # so dropout=0 still respects training-mode reqgrads — but model has dropout=0\n # Use cross-entropy on shifted next-token prediction\n out = model(calib_ids)\n logits = out.logits # [1, T, V]\n shift_logits = logits[:, :-1, :].contiguous().to(torch.float32)\n shift_labels = calib_ids[:, 1:].contiguous()\n loss = F.cross_entropy(shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1))\n log(f\"phase1: CE loss = {loss.item():.4f}\")\n loss.backward()\n\n importance = {}\n for L, g in gates.items():\n imp = g.alpha.grad.detach().abs().to(torch.float32).cpu()\n importance[L] = imp\n g.restore()\n model.eval()\n log(f\"phase1: importance computed for {len(importance)} layers\")\n return importance\n\n\n# ---------- phase 2: per-layer prune + lstsq refit ----------\n\n@torch.no_grad()\ndef prune_full_attention_inplace(layer, keep_q_heads, target_olin_input, model, calib_ids, device):\n \"\"\"\n Mask-prune full_attention layer:\n - Zero out q_proj rows for dropped Q heads (both q-half and gate-half).\n - Optionally drop matching KV heads if a whole GQA group is dropped.\n - lstsq-refit o_proj cols for kept Q heads on `target_olin_input`\n (the original o_proj input that we want to reproduce at this layer).\n\n target_olin_input: [B, T, n_q * head_dim] — tensor on cpu fp32.\n \"\"\"\n cfg = text_config(model)\n n_q = cfg.num_attention_heads\n n_kv = cfg.num_key_value_heads\n head_dim = cfg.head_dim\n group_size = n_q // n_kv # 16/4 = 4\n\n keep_q = sorted(int(i) for i in keep_q_heads)\n drop_q = [i for i in range(n_q) if i not in set(keep_q)]\n log(f\" full L?: keep_q={keep_q}, drop_q={drop_q}\")\n\n sa = layer.self_attn\n\n # Zero q_proj rows for dropped heads. q_proj is shape (n_q*head_dim*2, hidden).\n # The 2× factor: each head gets [q_h | gate_h]. Need to zero both halves.\n # Layout per head h of size 2*head_dim: rows [h*2*head_dim : (h+1)*2*head_dim]\n qw = sa.q_proj.weight.data # (n_q*head_dim*2, hidden)\n for h in drop_q:\n qw[h * 2 * head_dim : (h + 1) * 2 * head_dim, :] = 0\n # Optional KV pruning: drop whole groups\n keep_kv_groups = sorted({h // group_size for h in keep_q})\n drop_kv_groups = [g for g in range(n_kv) if g not in set(keep_kv_groups)]\n if drop_kv_groups:\n log(f\" full L?: drop KV groups {drop_kv_groups}\")\n for g in drop_kv_groups:\n sa.k_proj.weight.data[g * head_dim : (g + 1) * head_dim, :] = 0\n sa.v_proj.weight.data[g * head_dim : (g + 1) * head_dim, :] = 0\n\n # lstsq refit on o_proj\n # o_proj: y = W_O @ x where x = [B, T, n_q * head_dim], y = [B, T, hidden].\n # We want to find W_O' such that W_O' @ x' ≈ y_target, where x' has dropped-head\n # cols zeroed (since we zeroed q_proj rows above, those head cols of x are now 0).\n # But target was captured pre-prune. The refit uses the ACTUAL post-prune x,\n # which we get by re-running the layer's forward up to o_proj.\n #\n # Approach: re-forward the model with the zeroed q_proj, hook o_proj's input\n # at this layer, then solve lstsq(x_actual, y_target) for new W_O.\n # We do this in the caller (one re-fwd per layer is wasteful; instead we\n # re-forward once at the start of phase 2 after each prune).\n # Here we just receive x_actual via the caller after re-fwd.\n return # caller does the lstsq with re-captured x\n\n\n@torch.no_grad()\ndef prune_linear_attention_inplace(layer, keep_v_heads, model):\n \"\"\"\n Mask-prune linear_attention (gated DeltaNet):\n - Zero rows of in_proj_qkv corresponding to dropped V heads (value portion only).\n - Zero rows of in_proj_z, in_proj_b, in_proj_a for dropped V heads.\n - Zero conv1d depthwise channels for the value rows.\n - Zero entries of A_log, dt_bias for dropped V heads.\n - K-head dropping enabled when whole groups are pruned.\n The lstsq refit of out_proj is done by the caller after re-forward.\n \"\"\"\n cfg = text_config(model)\n n_v = cfg.linear_num_value_heads\n n_k = cfg.linear_num_key_heads\n head_v = cfg.linear_value_head_dim\n head_k = cfg.linear_key_head_dim\n key_dim = n_k * head_k\n value_dim = n_v * head_v\n\n keep_v = sorted(int(i) for i in keep_v_heads)\n drop_v = [i for i in range(n_v) if i not in set(keep_v)]\n log(f\" lin L?: keep_v={keep_v[:8]}{'...' if len(keep_v)>8 else ''}, drop_v={drop_v}\")\n\n la = layer.linear_attn\n\n # in_proj_qkv: (2*key_dim + value_dim, hidden). Layout: [Q | K | V] along output.\n # Value section starts at 2*key_dim and goes to 2*key_dim + value_dim.\n iqkv = la.in_proj_qkv.weight.data\n v_off = 2 * key_dim\n for h in drop_v:\n iqkv[v_off + h * head_v : v_off + (h + 1) * head_v, :] = 0\n # in_proj_z: (value_dim, hidden) — same layout as the V section\n iz = la.in_proj_z.weight.data\n for h in drop_v:\n iz[h * head_v : (h + 1) * head_v, :] = 0\n # in_proj_b, in_proj_a: (n_v, hidden)\n for h in drop_v:\n la.in_proj_b.weight.data[h, :] = 0\n la.in_proj_a.weight.data[h, :] = 0\n # conv1d: depthwise, (conv_dim, 1, kernel). conv_dim = 2*key_dim + value_dim\n cw = la.conv1d.weight.data # (conv_dim, 1, kernel)\n for h in drop_v:\n cw[v_off + h * head_v : v_off + (h + 1) * head_v, :, :] = 0\n # A_log, dt_bias: (n_v,)\n for h in drop_v:\n la.A_log.data[h] = 0 # log(1) = 0 means decay rate 1 — but unused since v rows are zeroed\n la.dt_bias.data[h] = 0\n\n # K-head group pruning if whole group dropped\n v_per_k = n_v // n_k # 32/16 = 2\n # Each k-head h_k corresponds to v-heads [h_k*v_per_k, (h_k+1)*v_per_k]\n drop_k_groups = []\n for h_k in range(n_k):\n v_in_group = list(range(h_k * v_per_k, (h_k + 1) * v_per_k))\n if all(v in drop_v for v in v_in_group):\n drop_k_groups.append(h_k)\n if drop_k_groups:\n log(f\" lin L?: drop K groups {drop_k_groups}\")\n for h_k in drop_k_groups:\n # K section in iqkv: [key_dim : 2*key_dim], offset 0 is Q\n for kk in range(2): # Q and K share dims layout [Q | K]\n base = kk * key_dim\n iqkv[base + h_k * head_k : base + (h_k + 1) * head_k, :] = 0\n cw[base + h_k * head_k : base + (h_k + 1) * head_k, :, :] = 0\n return\n\n\n@torch.no_grad()\ndef lstsq_refit_o_proj(proj: nn.Linear, x_actual: torch.Tensor, y_target: torch.Tensor, ridge: float = 1e-3):\n \"\"\"\n Solve W_new @ x_actual ≈ y_target via ridge-regularized least squares,\n write into proj.weight (shape: [out_features, in_features]).\n\n x_actual: [B, T, in_features] (post-prune o_proj input)\n y_target: [B, T, out_features] (pre-prune o_proj output target = W_orig @ x_orig)\n\n Note: y_target was captured as o_proj INPUT pre-prune. But we want post-proj\n output identity. Two options:\n (a) reproduce the post-projection output: y_target_out = W_orig @ x_orig_input.\n x_actual is the new o_proj input (with zeroed-head dims). lstsq finds W\n such that W @ x_actual ≈ W_orig @ x_orig_input.\n (b) reproduce the pre-prune o_proj input by adjusting kept cols: solve\n x_actual @ M ≈ x_orig_input. Not what we want.\n We use (a). The caller passes y_target = W_orig @ x_orig_input, computed\n by running the original projection on the captured input.\n \"\"\"\n assert x_actual.dim() == 3 and y_target.dim() == 3\n in_f = proj.in_features\n out_f = proj.out_features\n X = x_actual.reshape(-1, in_f).to(torch.float32)\n Y = y_target.reshape(-1, out_f).to(torch.float32)\n # Solve W (out_f x in_f): minimize ||X W^T - Y||^2 + ridge * ||W||^2\n # Normal equations: (X^T X + ridge I) W^T = X^T Y\n XtX = X.t() @ X\n XtY = X.t() @ Y\n XtX += ridge * torch.eye(in_f, dtype=XtX.dtype, device=XtX.device)\n W_t = torch.linalg.solve(XtX, XtY) # (in_f, out_f)\n W_new = W_t.t().contiguous() # (out_f, in_f)\n # Sanity: residual norm\n Y_hat = X @ W_t\n resid = (Y_hat - Y).pow(2).mean().sqrt().item()\n target_norm = Y.pow(2).mean().sqrt().item()\n rel = resid / max(target_norm, 1e-9)\n log(f\" lstsq refit: rel_resid={rel:.4f}, target_rms={target_norm:.4f}\")\n proj.weight.data.copy_(W_new.to(proj.weight.dtype).to(proj.weight.device))\n\n\n# ---------- main ----------\n\n@torch.no_grad()\ndef capture_proj_input_at_layer(model, calib_ids, layer_idx: int) -> torch.Tensor:\n \"\"\"Forward the (possibly modified) model and capture the o_proj/out_proj input at one specific layer.\"\"\"\n layer = get_layers(model)[layer_idx]\n if is_full(layer):\n proj = layer.self_attn.o_proj\n elif is_linear(layer):\n proj = layer.linear_attn.out_proj\n else:\n raise RuntimeError(\"layer has no attention\")\n\n captured = {}\n def hook(m, inp):\n captured[\"x\"] = inp[0].detach().to(torch.float32).cpu()\n h = proj.register_forward_pre_hook(hook)\n try:\n model(calib_ids)\n finally:\n h.remove()\n return captured[\"x\"]\n\n\ndef main():\n ap = argparse.ArgumentParser()\n ap.add_argument(\"--model-path\", required=True)\n ap.add_argument(\"--output\", required=True)\n ap.add_argument(\"--prune-full-frac\", type=float, default=0.0,\n help=\"fraction of Q heads to drop in full_attention layers\")\n ap.add_argument(\"--prune-linear-frac\", type=float, default=0.0,\n help=\"fraction of V heads to drop in linear_attention layers\")\n ap.add_argument(\"--calib-tokens\", type=int, default=2048)\n ap.add_argument(\"--device\", default=\"cuda\")\n ap.add_argument(\"--ridge\", type=float, default=1e-3)\n ap.add_argument(\"--smoke\", action=\"store_true\",\n help=\"only prune layer 3 (full) and/or layer 0 (linear) — sanity check\")\n args = ap.parse_args()\n\n log(f\"loading {args.model_path}\")\n tok = AutoTokenizer.from_pretrained(args.model_path)\n model = AutoModelForCausalLM.from_pretrained(args.model_path, torch_dtype=torch.bfloat16)\n model = model.to(args.device)\n\n # Disable use_cache: hooks easier and we never call generate during pruning\n model.config.use_cache = False\n if hasattr(model.config, \"text_config\"):\n model.config.text_config.use_cache = False\n model.eval()\n layers = get_layers(model)\n log(f\"model loaded: {len(layers)} layers; cfg full_heads={text_config(model).num_attention_heads} lin_v_heads={text_config(model).linear_num_value_heads}\")\n\n calib_ids = build_calib(tok, args.calib_tokens, args.device)\n\n # Phase 0: capture original o_proj/out_proj INPUT per layer (cpu fp32)\n targets = phase0_capture(model, calib_ids)\n # Convert each target to o_proj OUTPUT (the real target for healing)\n # by running it through the ORIGINAL projection weights, in fp32.\n log(\"phase0: computing pre-prune o_proj OUTPUTs as healing targets\")\n target_outputs = {}\n for L, x in targets.items():\n layer = layers[L]\n proj = layer.self_attn.o_proj if is_full(layer) else layer.linear_attn.out_proj\n W = proj.weight.data.to(torch.float32).cpu() # (out, in)\n # x: [B, T, in], y = x @ W^T: [B, T, out]\n y = x @ W.t()\n target_outputs[L] = y\n log(f\"phase0: {len(target_outputs)} target outputs prepared\")\n\n # Phase 1: importance\n importance = phase1_importance(model, calib_ids)\n\n # Phase 2: per-layer prune + lstsq refit, in forward order\n cfg = text_config(model)\n n_q_full = cfg.num_attention_heads\n n_v_lin = cfg.linear_num_value_heads\n\n smoke_layers = set()\n if args.smoke:\n # Pick first full layer and first linear layer\n for L, layer in enumerate(layers):\n if is_full(layer) and not any(is_full(layers[Li]) for Li in smoke_layers):\n smoke_layers.add(L)\n if is_linear(layer) and not any(is_linear(layers[Li]) for Li in smoke_layers):\n smoke_layers.add(L)\n if len(smoke_layers) >= 2:\n break\n log(f\"SMOKE mode: only pruning layers {sorted(smoke_layers)}\")\n\n n_drop_full = int(round(n_q_full * args.prune_full_frac))\n n_drop_lin = int(round(n_v_lin * args.prune_linear_frac))\n log(f\"phase2: drop {n_drop_full}/{n_q_full} full Q-heads per layer, {n_drop_lin}/{n_v_lin} linear V-heads per layer\")\n\n pruned_layers = []\n for L, layer in enumerate(layers):\n if args.smoke and L not in smoke_layers:\n continue\n if is_full(layer) and n_drop_full > 0:\n imp = importance[L]\n keep_idx = imp.argsort(descending=True)[: n_q_full - n_drop_full].tolist()\n log(f\"layer {L} [full]: importance min={imp.min():.3e} max={imp.max():.3e} mean={imp.mean():.3e}\")\n prune_full_attention_inplace(layer, keep_idx, None, model, calib_ids, args.device)\n # Re-capture o_proj input post-prune at this layer, then lstsq refit\n x_actual = capture_proj_input_at_layer(model, calib_ids, L)\n y_target = target_outputs[L]\n lstsq_refit_o_proj(layer.self_attn.o_proj, x_actual, y_target, ridge=args.ridge)\n pruned_layers.append((L, \"full\"))\n elif is_linear(layer) and n_drop_lin > 0:\n imp = importance[L]\n keep_idx = imp.argsort(descending=True)[: n_v_lin - n_drop_lin].tolist()\n log(f\"layer {L} [lin]: importance min={imp.min():.3e} max={imp.max():.3e} mean={imp.mean():.3e}\")\n prune_linear_attention_inplace(layer, keep_idx, model)\n x_actual = capture_proj_input_at_layer(model, calib_ids, L)\n y_target = target_outputs[L]\n lstsq_refit_o_proj(layer.linear_attn.out_proj, x_actual, y_target, ridge=args.ridge)\n pruned_layers.append((L, \"lin\"))\n\n # Final CE check\n with torch.no_grad():\n out = model(calib_ids)\n logits = out.logits\n shift_logits = logits[:, :-1, :].contiguous().to(torch.float32)\n shift_labels = calib_ids[:, 1:].contiguous()\n ce_post = F.cross_entropy(shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1)).item()\n log(f\"final CE on calibration: {ce_post:.4f}\")\n\n # Save\n out_dir = Path(args.output)\n out_dir.mkdir(parents=True, exist_ok=True)\n log(f\"saving to {out_dir}\")\n model.save_pretrained(out_dir)\n tok.save_pretrained(out_dir)\n\n # Drop a manifest of what we did\n manifest = {\n \"source\": str(args.model_path),\n \"prune_full_frac\": args.prune_full_frac,\n \"prune_linear_frac\": args.prune_linear_frac,\n \"n_drop_full\": n_drop_full,\n \"n_drop_linear\": n_drop_lin,\n \"calib_tokens\": args.calib_tokens,\n \"ridge\": args.ridge,\n \"smoke\": args.smoke,\n \"pruned_layers\": pruned_layers,\n \"final_ce\": ce_post,\n }\n (out_dir / \"prune_manifest.json\").write_text(json.dumps(manifest, indent=2))\n log(\"done\")\n\n\nif __name__ == \"__main__\":\n main()\n"},"tool_response":{"_truncated":"{\"type\":\"create\",\"filePath\":\"/shared/dev/omnimergekit/recipes/microcoder_4b/prune_local_heal.py\",\"content\":\"#!/usr/bin/env python3\\n\\\"\\\"\\\"\\nprune_local_heal.py — Qwen3.5 hybrid-arch attention head pruning with\\n backprop-free local healing (W_O lstsq refit).\\n\\nTargets the Qwen3.5 hybrid architecture which mixes:\\n - full_attention layers (Qwen3_5Attention with gated query, GQA 16:4, head_dim=256)\\n - linear_attention layers (Qwen3_5GatedDeltaNet, 32 v_heads / 16 k_heads, head_dim=128)\\n\\nTwo prune levers:\\n --prune-full-frac : fraction of Q heads to drop in full_attention layers\\n --prune-linear-frac : fraction of V heads to drop in linear_attention layers\\n\\nPhases:\\n 0. Capture per-layer o_proj-input (full_attn) and out_proj-input (linear_attn)\\n on a calibration set — these are the targets for healing.\\n 1. Score per-head importance via Michel-style gradient on per-head scale α=1.\\n 2. For each layer, in forward order, on the *current* (already-modified) model:\\n a. Re-capture pre-proj input\\n b. Zero rows for dropped heads in q/k/v / in_proj_qkv (and corresponding\\n rows in z/b/a/conv1d/A_log/dt_bias for linear_attn)\\n c. lstsq refit the kept columns of o_proj / out_proj so the post-proj\\n output matches the original target\\n\\nThe model stays full-size on disk (mask-style); a separate physical-resize\\nstep converts to a smaller model after eval validation. Mask-pruning is what\\nlstsq refit operates on, and it preserves output quality identically — only\\ninference compute differs.\\n\\\"\\\"\\\"\\nfrom __future__ import annotations\\n\\nimport argparse\\nimport json\\nimport os\\nimport time\\nfrom pathlib import Path\\n\\nimport torch\\nimport torch.nn as nn\\nimport torch.nn.functional as F\\nfrom transformers import AutoModelForCausalLM, AutoTokenizer\\n\\n\\ndef log(msg: str):\\n print(f\\\"[{time.strftime('%H:%M:%S')}] {msg}\\\", flush=True)\\n\\n\\n# ---------- calibration ----------\\n\\nCALIB_TEXTS = [\\n # Co"},"tool_use_id":"toolu_01HVaYqJUYe7SbK3zE8iSdue","cwd":"/srv/dev-disk-by-uuid-f8b1803e-334f-4f4b-af3b-f802bb6883c5/backup_models/4b_phase1/merged/v2i-jv-base-task-arith"}
6
+ {"timestamp":"2026-05-04T07:10:10.273Z","session_id":"4208dc59-08bc-4b1e-8d4e-8ec3407fd4ac","hook_event_name":"PostToolUse","tool_name":"Edit","tool_input":{"file_path":"/shared/dev/omnimergekit/recipes/microcoder_4b/prune_local_heal.py","old_string":"import argparse\nimport json\nimport os\nimport time","new_string":"import argparse\nimport json\nimport time","replace_all":false},"tool_response":{"_truncated":"{\"filePath\":\"/shared/dev/omnimergekit/recipes/microcoder_4b/prune_local_heal.py\",\"oldString\":\"import argparse\\nimport json\\nimport os\\nimport time\",\"newString\":\"import argparse\\nimport json\\nimport time\",\"originalFile\":\"#!/usr/bin/env python3\\n\\\"\\\"\\\"\\nprune_local_heal.py — Qwen3.5 hybrid-arch attention head pruning with\\n backprop-free local healing (W_O lstsq refit).\\n\\nTargets the Qwen3.5 hybrid architecture which mixes:\\n - full_attention layers (Qwen3_5Attention with gated query, GQA 16:4, head_dim=256)\\n - linear_attention layers (Qwen3_5GatedDeltaNet, 32 v_heads / 16 k_heads, head_dim=128)\\n\\nTwo prune levers:\\n --prune-full-frac : fraction of Q heads to drop in full_attention layers\\n --prune-linear-frac : fraction of V heads to drop in linear_attention layers\\n\\nPhases:\\n 0. Capture per-layer o_proj-input (full_attn) and out_proj-input (linear_attn)\\n on a calibration set — these are the targets for healing.\\n 1. Score per-head importance via Michel-style gradient on per-head scale α=1.\\n 2. For each layer, in forward order, on the *current* (already-modified) model:\\n a. Re-capture pre-proj input\\n b. Zero rows for dropped heads in q/k/v / in_proj_qkv (and corresponding\\n rows in z/b/a/conv1d/A_log/dt_bias for linear_attn)\\n c. lstsq refit the kept columns of o_proj / out_proj so the post-proj\\n output matches the original target\\n\\nThe model stays full-size on disk (mask-style); a separate physical-resize\\nstep converts to a smaller model after eval validation. Mask-pruning is what\\nlstsq refit operates on, and it preserves output quality identically — only\\ninference compute differs.\\n\\\"\\\"\\\"\\nfrom __future__ import annotations\\n\\nimport argparse\\nimport json\\nimport os\\nimport time\\nfrom pathlib import Path\\n\\nimport torch\\nimport torch.nn as nn\\nimport torch.nn.functional as F\\nfrom transformers import AutoModelForCausalLM, AutoTokenizer\\n\\n\\ndef log(msg: str):\\n print(f\\\"[{time.s"},"tool_use_id":"toolu_01WasHPd6eJF6aCAPkUAhTSw","cwd":"/srv/dev-disk-by-uuid-f8b1803e-334f-4f4b-af3b-f802bb6883c5/backup_models/4b_phase1/merged/v2i-jv-base-task-arith"}
7
+ {"timestamp":"2026-05-04T07:10:15.622Z","session_id":"4208dc59-08bc-4b1e-8d4e-8ec3407fd4ac","hook_event_name":"PostToolUse","tool_name":"Edit","tool_input":{"file_path":"/shared/dev/omnimergekit/recipes/microcoder_4b/prune_local_heal.py","old_string":" key_dim = n_k * head_k\n value_dim = n_v * head_v\n\n keep_v = sorted(int(i) for i in keep_v_heads)","new_string":" key_dim = n_k * head_k\n\n keep_v = sorted(int(i) for i in keep_v_heads)","replace_all":false},"tool_response":{"_truncated":"{\"filePath\":\"/shared/dev/omnimergekit/recipes/microcoder_4b/prune_local_heal.py\",\"oldString\":\" key_dim = n_k * head_k\\n value_dim = n_v * head_v\\n\\n keep_v = sorted(int(i) for i in keep_v_heads)\",\"newString\":\" key_dim = n_k * head_k\\n\\n keep_v = sorted(int(i) for i in keep_v_heads)\",\"originalFile\":\"#!/usr/bin/env python3\\n\\\"\\\"\\\"\\nprune_local_heal.py — Qwen3.5 hybrid-arch attention head pruning with\\n backprop-free local healing (W_O lstsq refit).\\n\\nTargets the Qwen3.5 hybrid architecture which mixes:\\n - full_attention layers (Qwen3_5Attention with gated query, GQA 16:4, head_dim=256)\\n - linear_attention layers (Qwen3_5GatedDeltaNet, 32 v_heads / 16 k_heads, head_dim=128)\\n\\nTwo prune levers:\\n --prune-full-frac : fraction of Q heads to drop in full_attention layers\\n --prune-linear-frac : fraction of V heads to drop in linear_attention layers\\n\\nPhases:\\n 0. Capture per-layer o_proj-input (full_attn) and out_proj-input (linear_attn)\\n on a calibration set — these are the targets for healing.\\n 1. Score per-head importance via Michel-style gradient on per-head scale α=1.\\n 2. For each layer, in forward order, on the *current* (already-modified) model:\\n a. Re-capture pre-proj input\\n b. Zero rows for dropped heads in q/k/v / in_proj_qkv (and corresponding\\n rows in z/b/a/conv1d/A_log/dt_bias for linear_attn)\\n c. lstsq refit the kept columns of o_proj / out_proj so the post-proj\\n output matches the original target\\n\\nThe model stays full-size on disk (mask-style); a separate physical-resize\\nstep converts to a smaller model after eval validation. Mask-pruning is what\\nlstsq refit operates on, and it preserves output quality identically — only\\ninference compute differs.\\n\\\"\\\"\\\"\\nfrom __future__ import annotations\\n\\nimport argparse\\nimport json\\nimport time\\nfrom pathlib import Path\\n\\nimport torch\\nimport torch.nn as nn\\nimport torch.nn.functional as F\\nfrom transformers impor"},"tool_use_id":"toolu_01Eo49sKnQyNLedwvVR8Ygqk","cwd":"/srv/dev-disk-by-uuid-f8b1803e-334f-4f4b-af3b-f802bb6883c5/backup_models/4b_phase1/merged/v2i-jv-base-task-arith"}
.caliber/learning/state.json ADDED
@@ -0,0 +1,6 @@
 
 
 
 
 
 
 
1
+ {
2
+ "sessionId": "4208dc59-08bc-4b1e-8d4e-8ec3407fd4ac",
3
+ "eventCount": 7,
4
+ "lastAnalysisTimestamp": null,
5
+ "lastAnalysisEventCount": 0
6
+ }
.gitattributes CHANGED
@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
33
  *.zip filter=lfs diff=lfs merge=lfs -text
34
  *.zst filter=lfs diff=lfs merge=lfs -text
35
  *tfevents* filter=lfs diff=lfs merge=lfs -text
 
 
33
  *.zip filter=lfs diff=lfs merge=lfs -text
34
  *.zst filter=lfs diff=lfs merge=lfs -text
35
  *tfevents* filter=lfs diff=lfs merge=lfs -text
36
+ tokenizer.json filter=lfs diff=lfs merge=lfs -text
chat_template.jinja ADDED
@@ -0,0 +1,88 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {%- if tools %}
2
+ {{- '<|im_start|>system\n' }}
3
+ {%- if messages[0].role == 'system' %}
4
+ {{- messages[0].content + '\n\n' }}
5
+ {%- endif %}
6
+ {{- "# Tools\n\nYou may call one or more functions to assist with the user query.\n\nYou are provided with function signatures within <tools></tools> XML tags:\n<tools>" }}
7
+ {%- for tool in tools %}
8
+ {{- "\n" }}
9
+ {{- tool | tojson }}
10
+ {%- endfor %}
11
+ {{- "\n</tools>\n\nFor each function call, return a json object with function name and arguments within <tool_call></tool_call> XML tags:\n<tool_call>\n{\"name\": <function-name>, \"arguments\": <args-json-object>}\n</tool_call><|im_end|>\n" }}
12
+ {%- else %}
13
+ {%- if messages[0].role == 'system' %}
14
+ {{- '<|im_start|>system\n' + messages[0].content + '<|im_end|>\n' }}
15
+ {%- endif %}
16
+ {%- endif %}
17
+ {%- set ns = namespace(multi_step_tool=true, last_query_index=messages|length - 1) %}
18
+ {%- for message in messages[::-1] %}
19
+ {%- set index = (messages|length - 1) - loop.index0 %}
20
+ {%- if ns.multi_step_tool and message.role == "user" and message.content is string and not(message.content.startswith('<tool_response>') and message.content.endswith('</tool_response>')) %}
21
+ {%- set ns.multi_step_tool = false %}
22
+ {%- set ns.last_query_index = index %}
23
+ {%- endif %}
24
+ {%- endfor %}
25
+ {%- for message in messages %}
26
+ {%- if message.content is string %}
27
+ {%- set content = message.content %}
28
+ {%- else %}
29
+ {%- set content = '' %}
30
+ {%- endif %}
31
+ {%- if (message.role == "user") or (message.role == "system" and not loop.first) %}
32
+ {{- '<|im_start|>' + message.role + '\n' + content + '<|im_end|>' + '\n' }}
33
+ {%- elif message.role == "assistant" %}
34
+ {%- set reasoning_content = '' %}
35
+ {%- if message.reasoning_content is string %}
36
+ {%- set reasoning_content = message.reasoning_content %}
37
+ {%- else %}
38
+ {%- if '</think>' in content %}
39
+ {%- set reasoning_content = content.split('</think>')[0].rstrip('\n').split('<think>')[-1].lstrip('\n') %}
40
+ {%- set content = content.split('</think>')[-1].lstrip('\n') %}
41
+ {%- endif %}
42
+ {%- endif %}
43
+ {%- if loop.index0 > ns.last_query_index %}
44
+ {%- if loop.last or (not loop.last and reasoning_content) %}
45
+ {{- '<|im_start|>' + message.role + '\n<think>\n' + reasoning_content.strip('\n') + '\n</think>\n\n' + content.lstrip('\n') }}
46
+ {%- else %}
47
+ {{- '<|im_start|>' + message.role + '\n' + content }}
48
+ {%- endif %}
49
+ {%- else %}
50
+ {{- '<|im_start|>' + message.role + '\n' + content }}
51
+ {%- endif %}
52
+ {%- if message.tool_calls %}
53
+ {%- for tool_call in message.tool_calls %}
54
+ {%- if (loop.first and content) or (not loop.first) %}
55
+ {{- '\n' }}
56
+ {%- endif %}
57
+ {%- if tool_call.function %}
58
+ {%- set tool_call = tool_call.function %}
59
+ {%- endif %}
60
+ {{- '<tool_call>\n{"name": "' }}
61
+ {{- tool_call.name }}
62
+ {{- '", "arguments": ' }}
63
+ {%- if tool_call.arguments is string %}
64
+ {{- tool_call.arguments }}
65
+ {%- else %}
66
+ {{- tool_call.arguments | tojson }}
67
+ {%- endif %}
68
+ {{- '}\n</tool_call>' }}
69
+ {%- endfor %}
70
+ {%- endif %}
71
+ {{- '<|im_end|>\n' }}
72
+ {%- elif message.role == "tool" %}
73
+ {%- if loop.first or (messages[loop.index0 - 1].role != "tool") %}
74
+ {{- '<|im_start|>user' }}
75
+ {%- endif %}
76
+ {{- '\n<tool_response>\n' }}
77
+ {{- content }}
78
+ {{- '\n</tool_response>' }}
79
+ {%- if loop.last or (messages[loop.index0 + 1].role != "tool") %}
80
+ {{- '<|im_end|>\n' }}
81
+ {%- endif %}
82
+ {%- endif %}
83
+ {%- endfor %}
84
+ {%- if add_generation_prompt %}
85
+ {{- '<|im_start|>assistant
86
+ <think>
87
+ ' }}
88
+ {%- endif %}
config.json ADDED
@@ -0,0 +1,114 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "architectures": [
3
+ "Qwen3_5ForConditionalGeneration"
4
+ ],
5
+ "torch_dtype": "bfloat16",
6
+ "eos_token_id": 248046,
7
+ "image_token_id": 248056,
8
+ "model_name": "unsloth/Qwen3.5-4B",
9
+ "model_type": "qwen3_5",
10
+ "pad_token_id": 248055,
11
+ "text_config": {
12
+ "attention_bias": false,
13
+ "attention_dropout": 0.0,
14
+ "attn_output_gate": true,
15
+ "bos_token_id": null,
16
+ "torch_dtype": "bfloat16",
17
+ "eos_token_id": 248044,
18
+ "full_attention_interval": 4,
19
+ "head_dim": 256,
20
+ "hidden_act": "silu",
21
+ "hidden_size": 2560,
22
+ "initializer_range": 0.02,
23
+ "intermediate_size": 9216,
24
+ "layer_types": [
25
+ "linear_attention",
26
+ "linear_attention",
27
+ "linear_attention",
28
+ "full_attention",
29
+ "linear_attention",
30
+ "linear_attention",
31
+ "linear_attention",
32
+ "full_attention",
33
+ "linear_attention",
34
+ "linear_attention",
35
+ "linear_attention",
36
+ "full_attention",
37
+ "linear_attention",
38
+ "linear_attention",
39
+ "linear_attention",
40
+ "full_attention",
41
+ "linear_attention",
42
+ "linear_attention",
43
+ "linear_attention",
44
+ "full_attention",
45
+ "linear_attention",
46
+ "linear_attention",
47
+ "linear_attention",
48
+ "full_attention",
49
+ "linear_attention",
50
+ "linear_attention",
51
+ "linear_attention",
52
+ "full_attention",
53
+ "linear_attention",
54
+ "linear_attention",
55
+ "linear_attention",
56
+ "full_attention"
57
+ ],
58
+ "linear_conv_kernel_dim": 4,
59
+ "linear_key_head_dim": 128,
60
+ "linear_num_key_heads": 16,
61
+ "linear_num_value_heads": 32,
62
+ "linear_value_head_dim": 128,
63
+ "mamba_ssm_dtype": "float32",
64
+ "max_position_embeddings": 262144,
65
+ "mlp_only_layers": [],
66
+ "model_type": "qwen3_5_text",
67
+ "mtp_num_hidden_layers": 1,
68
+ "mtp_use_dedicated_embeddings": false,
69
+ "num_attention_heads": 16,
70
+ "num_hidden_layers": 32,
71
+ "num_key_value_heads": 4,
72
+ "pad_token_id": null,
73
+ "partial_rotary_factor": 0.25,
74
+ "rms_norm_eps": 1e-06,
75
+ "rope_parameters": {
76
+ "mrope_interleaved": true,
77
+ "mrope_section": [
78
+ 11,
79
+ 11,
80
+ 10
81
+ ],
82
+ "partial_rotary_factor": 0.25,
83
+ "rope_theta": 10000000,
84
+ "rope_type": "default"
85
+ },
86
+ "tie_word_embeddings": true,
87
+ "use_cache": true,
88
+ "vocab_size": 248320
89
+ },
90
+ "tie_word_embeddings": true,
91
+ "unsloth_fixed": true,
92
+ "unsloth_version": "2026.3.4",
93
+ "use_cache": false,
94
+ "video_token_id": 248057,
95
+ "vision_config": {
96
+ "deepstack_visual_indexes": [],
97
+ "depth": 24,
98
+ "torch_dtype": "bfloat16",
99
+ "hidden_act": "gelu_pytorch_tanh",
100
+ "hidden_size": 1024,
101
+ "in_channels": 3,
102
+ "initializer_range": 0.02,
103
+ "intermediate_size": 4096,
104
+ "model_type": "qwen3_5",
105
+ "num_heads": 16,
106
+ "num_position_embeddings": 2304,
107
+ "out_hidden_size": 2560,
108
+ "patch_size": 16,
109
+ "spatial_merge_size": 2,
110
+ "temporal_patch_size": 2
111
+ },
112
+ "vision_end_token_id": 248054,
113
+ "vision_start_token_id": 248053
114
+ }
model-00001-of-00002.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:45d2fec62c40a19c4b439671b0abaa55cec317bd39e6ee32760507a77e041d61
3
+ size 5376584768
model-00002-of-00002.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:15de8b5eb052e291abb1fdf903642224fc9270417b0ee18025e4a902a150cb62
3
+ size 3943243384
model.safetensors.index.json ADDED
@@ -0,0 +1,745 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "metadata": {
3
+ "total_size": 9319828152
4
+ },
5
+ "weight_map": {
6
+ "model.language_model.embed_tokens.weight": "model-00001-of-00002.safetensors",
7
+ "model.language_model.layers.1.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
8
+ "model.language_model.layers.1.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
9
+ "model.language_model.layers.1.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
10
+ "model.language_model.layers.10.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
11
+ "model.language_model.layers.10.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
12
+ "model.language_model.layers.10.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
13
+ "model.language_model.layers.11.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
14
+ "model.language_model.layers.11.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
15
+ "model.language_model.layers.11.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
16
+ "model.language_model.layers.18.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
17
+ "model.language_model.layers.18.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
18
+ "model.language_model.layers.18.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
19
+ "model.language_model.layers.19.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
20
+ "model.language_model.layers.19.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
21
+ "model.language_model.layers.19.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
22
+ "model.language_model.layers.12.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
23
+ "model.language_model.layers.12.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
24
+ "model.language_model.layers.12.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
25
+ "model.language_model.layers.13.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
26
+ "model.language_model.layers.13.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
27
+ "model.language_model.layers.13.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
28
+ "model.language_model.layers.7.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
29
+ "model.language_model.layers.8.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
30
+ "model.language_model.layers.8.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
31
+ "model.language_model.layers.8.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
32
+ "model.language_model.layers.21.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
33
+ "model.language_model.layers.21.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
34
+ "model.language_model.layers.21.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
35
+ "model.language_model.layers.22.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
36
+ "model.language_model.layers.22.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
37
+ "model.language_model.layers.22.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
38
+ "model.language_model.layers.23.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
39
+ "model.language_model.layers.23.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
40
+ "model.language_model.layers.23.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
41
+ "model.language_model.layers.31.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
42
+ "model.language_model.layers.31.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
43
+ "model.language_model.layers.31.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
44
+ "model.language_model.layers.4.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
45
+ "model.language_model.layers.4.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
46
+ "model.language_model.layers.4.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
47
+ "model.language_model.layers.14.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
48
+ "model.language_model.layers.14.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
49
+ "model.language_model.layers.14.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
50
+ "model.language_model.layers.15.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
51
+ "model.language_model.layers.15.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
52
+ "model.language_model.layers.15.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
53
+ "model.language_model.layers.9.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
54
+ "model.language_model.layers.9.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
55
+ "model.language_model.layers.9.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
56
+ "mtp.layers.0.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
57
+ "mtp.layers.0.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
58
+ "mtp.layers.0.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
59
+ "model.language_model.layers.29.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
60
+ "model.language_model.layers.3.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
61
+ "model.language_model.layers.3.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
62
+ "model.language_model.layers.3.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
63
+ "model.language_model.layers.30.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
64
+ "model.language_model.layers.30.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
65
+ "model.language_model.layers.30.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
66
+ "model.language_model.layers.16.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
67
+ "model.language_model.layers.16.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
68
+ "model.language_model.layers.16.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
69
+ "model.language_model.layers.17.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
70
+ "model.language_model.layers.17.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
71
+ "model.language_model.layers.17.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
72
+ "model.language_model.layers.24.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
73
+ "model.language_model.layers.24.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
74
+ "model.language_model.layers.24.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
75
+ "model.language_model.layers.5.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
76
+ "model.language_model.layers.5.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
77
+ "model.language_model.layers.5.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
78
+ "model.language_model.layers.6.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
79
+ "model.language_model.layers.6.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
80
+ "model.language_model.layers.6.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
81
+ "model.language_model.layers.7.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
82
+ "model.language_model.layers.7.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
83
+ "model.language_model.layers.2.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
84
+ "model.language_model.layers.2.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
85
+ "model.language_model.layers.2.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
86
+ "model.language_model.layers.20.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
87
+ "model.language_model.layers.20.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
88
+ "model.language_model.layers.20.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
89
+ "model.language_model.layers.25.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
90
+ "model.language_model.layers.25.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
91
+ "model.language_model.layers.25.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
92
+ "model.language_model.layers.26.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
93
+ "model.language_model.layers.26.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
94
+ "model.language_model.layers.26.mlp.up_proj.weight": "model-00002-of-00002.safetensors",
95
+ "model.language_model.layers.27.mlp.down_proj.weight": "model-00002-of-00002.safetensors",
96
+ "model.language_model.layers.27.mlp.gate_proj.weight": "model-00002-of-00002.safetensors",
97
+ "model.language_model.layers.27.mlp.up_proj.weight": "model-00002-of-00002.safetensors",
98
+ "model.language_model.layers.28.mlp.down_proj.weight": "model-00002-of-00002.safetensors",
99
+ "model.language_model.layers.28.mlp.gate_proj.weight": "model-00002-of-00002.safetensors",
100
+ "model.language_model.layers.28.mlp.up_proj.weight": "model-00002-of-00002.safetensors",
101
+ "model.language_model.layers.29.mlp.gate_proj.weight": "model-00002-of-00002.safetensors",
102
+ "model.language_model.layers.29.mlp.up_proj.weight": "model-00002-of-00002.safetensors",
103
+ "model.language_model.layers.0.mlp.down_proj.weight": "model-00002-of-00002.safetensors",
104
+ "model.language_model.layers.0.mlp.gate_proj.weight": "model-00002-of-00002.safetensors",
105
+ "model.language_model.layers.0.mlp.up_proj.weight": "model-00002-of-00002.safetensors",
106
+ "model.language_model.layers.10.linear_attn.in_proj_qkv.weight": "model-00002-of-00002.safetensors",
107
+ "model.language_model.layers.11.self_attn.q_proj.weight": "model-00002-of-00002.safetensors",
108
+ "model.language_model.layers.18.linear_attn.in_proj_qkv.weight": "model-00002-of-00002.safetensors",
109
+ "model.language_model.layers.19.self_attn.q_proj.weight": "model-00002-of-00002.safetensors",
110
+ "model.language_model.layers.2.linear_attn.in_proj_qkv.weight": "model-00002-of-00002.safetensors",
111
+ "model.language_model.layers.12.linear_attn.in_proj_qkv.weight": "model-00002-of-00002.safetensors",
112
+ "model.language_model.layers.13.linear_attn.in_proj_qkv.weight": "model-00002-of-00002.safetensors",
113
+ "model.language_model.layers.7.self_attn.q_proj.weight": "model-00002-of-00002.safetensors",
114
+ "model.language_model.layers.8.linear_attn.in_proj_qkv.weight": "model-00002-of-00002.safetensors",
115
+ "model.language_model.layers.22.linear_attn.in_proj_qkv.weight": "model-00002-of-00002.safetensors",
116
+ "model.language_model.layers.23.self_attn.q_proj.weight": "model-00002-of-00002.safetensors",
117
+ "model.language_model.layers.31.self_attn.q_proj.weight": "model-00002-of-00002.safetensors",
118
+ "model.language_model.layers.4.linear_attn.in_proj_qkv.weight": "model-00002-of-00002.safetensors",
119
+ "model.language_model.layers.14.linear_attn.in_proj_qkv.weight": "model-00002-of-00002.safetensors",
120
+ "model.language_model.layers.15.self_attn.q_proj.weight": "model-00002-of-00002.safetensors",
121
+ "model.language_model.layers.9.linear_attn.in_proj_qkv.weight": "model-00002-of-00002.safetensors",
122
+ "mtp.layers.0.self_attn.q_proj.weight": "model-00002-of-00002.safetensors",
123
+ "model.language_model.layers.3.self_attn.q_proj.weight": "model-00002-of-00002.safetensors",
124
+ "model.language_model.layers.30.linear_attn.in_proj_qkv.weight": "model-00002-of-00002.safetensors",
125
+ "model.language_model.layers.16.linear_attn.in_proj_qkv.weight": "model-00002-of-00002.safetensors",
126
+ "model.language_model.layers.17.linear_attn.in_proj_qkv.weight": "model-00002-of-00002.safetensors",
127
+ "model.language_model.layers.24.linear_attn.in_proj_qkv.weight": "model-00002-of-00002.safetensors",
128
+ "model.language_model.layers.25.linear_attn.in_proj_qkv.weight": "model-00002-of-00002.safetensors",
129
+ "model.language_model.layers.5.linear_attn.in_proj_qkv.weight": "model-00002-of-00002.safetensors",
130
+ "model.language_model.layers.6.linear_attn.in_proj_qkv.weight": "model-00002-of-00002.safetensors",
131
+ "model.language_model.layers.20.linear_attn.in_proj_qkv.weight": "model-00002-of-00002.safetensors",
132
+ "model.language_model.layers.21.linear_attn.in_proj_qkv.weight": "model-00002-of-00002.safetensors",
133
+ "model.language_model.layers.26.linear_attn.in_proj_qkv.weight": "model-00002-of-00002.safetensors",
134
+ "model.language_model.layers.27.self_attn.q_proj.weight": "model-00002-of-00002.safetensors",
135
+ "model.language_model.layers.28.linear_attn.in_proj_qkv.weight": "model-00002-of-00002.safetensors",
136
+ "model.language_model.layers.29.linear_attn.in_proj_qkv.weight": "model-00002-of-00002.safetensors",
137
+ "model.language_model.layers.0.linear_attn.in_proj_qkv.weight": "model-00002-of-00002.safetensors",
138
+ "model.language_model.layers.1.linear_attn.in_proj_qkv.weight": "model-00002-of-00002.safetensors",
139
+ "model.visual.merger.linear_fc1.weight": "model-00002-of-00002.safetensors",
140
+ "mtp.fc.weight": "model-00002-of-00002.safetensors",
141
+ "model.language_model.layers.10.linear_attn.in_proj_z.weight": "model-00002-of-00002.safetensors",
142
+ "model.language_model.layers.10.linear_attn.out_proj.weight": "model-00002-of-00002.safetensors",
143
+ "model.language_model.layers.11.self_attn.o_proj.weight": "model-00002-of-00002.safetensors",
144
+ "model.language_model.layers.18.linear_attn.in_proj_z.weight": "model-00002-of-00002.safetensors",
145
+ "model.language_model.layers.18.linear_attn.out_proj.weight": "model-00002-of-00002.safetensors",
146
+ "model.language_model.layers.19.self_attn.o_proj.weight": "model-00002-of-00002.safetensors",
147
+ "model.language_model.layers.2.linear_attn.in_proj_z.weight": "model-00002-of-00002.safetensors",
148
+ "model.language_model.layers.12.linear_attn.in_proj_z.weight": "model-00002-of-00002.safetensors",
149
+ "model.language_model.layers.12.linear_attn.out_proj.weight": "model-00002-of-00002.safetensors",
150
+ "model.language_model.layers.13.linear_attn.in_proj_z.weight": "model-00002-of-00002.safetensors",
151
+ "model.language_model.layers.13.linear_attn.out_proj.weight": "model-00002-of-00002.safetensors",
152
+ "model.language_model.layers.7.self_attn.o_proj.weight": "model-00002-of-00002.safetensors",
153
+ "model.language_model.layers.8.linear_attn.in_proj_z.weight": "model-00002-of-00002.safetensors",
154
+ "model.language_model.layers.8.linear_attn.out_proj.weight": "model-00002-of-00002.safetensors",
155
+ "model.language_model.layers.21.linear_attn.out_proj.weight": "model-00002-of-00002.safetensors",
156
+ "model.language_model.layers.22.linear_attn.in_proj_z.weight": "model-00002-of-00002.safetensors",
157
+ "model.language_model.layers.22.linear_attn.out_proj.weight": "model-00002-of-00002.safetensors",
158
+ "model.language_model.layers.23.self_attn.o_proj.weight": "model-00002-of-00002.safetensors",
159
+ "model.language_model.layers.31.self_attn.o_proj.weight": "model-00002-of-00002.safetensors",
160
+ "model.language_model.layers.4.linear_attn.in_proj_z.weight": "model-00002-of-00002.safetensors",
161
+ "model.language_model.layers.4.linear_attn.out_proj.weight": "model-00002-of-00002.safetensors",
162
+ "model.language_model.layers.14.linear_attn.in_proj_z.weight": "model-00002-of-00002.safetensors",
163
+ "model.language_model.layers.14.linear_attn.out_proj.weight": "model-00002-of-00002.safetensors",
164
+ "model.language_model.layers.15.self_attn.o_proj.weight": "model-00002-of-00002.safetensors",
165
+ "model.language_model.layers.9.linear_attn.in_proj_z.weight": "model-00002-of-00002.safetensors",
166
+ "model.language_model.layers.9.linear_attn.out_proj.weight": "model-00002-of-00002.safetensors",
167
+ "mtp.layers.0.self_attn.o_proj.weight": "model-00002-of-00002.safetensors",
168
+ "model.language_model.layers.3.self_attn.o_proj.weight": "model-00002-of-00002.safetensors",
169
+ "model.language_model.layers.30.linear_attn.in_proj_z.weight": "model-00002-of-00002.safetensors",
170
+ "model.language_model.layers.30.linear_attn.out_proj.weight": "model-00002-of-00002.safetensors",
171
+ "model.language_model.layers.16.linear_attn.in_proj_z.weight": "model-00002-of-00002.safetensors",
172
+ "model.language_model.layers.16.linear_attn.out_proj.weight": "model-00002-of-00002.safetensors",
173
+ "model.language_model.layers.17.linear_attn.in_proj_z.weight": "model-00002-of-00002.safetensors",
174
+ "model.language_model.layers.17.linear_attn.out_proj.weight": "model-00002-of-00002.safetensors",
175
+ "model.language_model.layers.24.linear_attn.in_proj_z.weight": "model-00002-of-00002.safetensors",
176
+ "model.language_model.layers.24.linear_attn.out_proj.weight": "model-00002-of-00002.safetensors",
177
+ "model.language_model.layers.25.linear_attn.in_proj_z.weight": "model-00002-of-00002.safetensors",
178
+ "model.language_model.layers.25.linear_attn.out_proj.weight": "model-00002-of-00002.safetensors",
179
+ "model.language_model.layers.5.linear_attn.in_proj_z.weight": "model-00002-of-00002.safetensors",
180
+ "model.language_model.layers.5.linear_attn.out_proj.weight": "model-00002-of-00002.safetensors",
181
+ "model.language_model.layers.6.linear_attn.in_proj_z.weight": "model-00002-of-00002.safetensors",
182
+ "model.language_model.layers.6.linear_attn.out_proj.weight": "model-00002-of-00002.safetensors",
183
+ "model.language_model.layers.2.linear_attn.out_proj.weight": "model-00002-of-00002.safetensors",
184
+ "model.language_model.layers.20.linear_attn.in_proj_z.weight": "model-00002-of-00002.safetensors",
185
+ "model.language_model.layers.20.linear_attn.out_proj.weight": "model-00002-of-00002.safetensors",
186
+ "model.language_model.layers.21.linear_attn.in_proj_z.weight": "model-00002-of-00002.safetensors",
187
+ "model.language_model.layers.26.linear_attn.in_proj_z.weight": "model-00002-of-00002.safetensors",
188
+ "model.language_model.layers.26.linear_attn.out_proj.weight": "model-00002-of-00002.safetensors",
189
+ "model.language_model.layers.27.self_attn.o_proj.weight": "model-00002-of-00002.safetensors",
190
+ "model.language_model.layers.28.linear_attn.in_proj_z.weight": "model-00002-of-00002.safetensors",
191
+ "model.language_model.layers.28.linear_attn.out_proj.weight": "model-00002-of-00002.safetensors",
192
+ "model.language_model.layers.29.linear_attn.in_proj_z.weight": "model-00002-of-00002.safetensors",
193
+ "model.language_model.layers.29.linear_attn.out_proj.weight": "model-00002-of-00002.safetensors",
194
+ "model.language_model.layers.0.linear_attn.in_proj_z.weight": "model-00002-of-00002.safetensors",
195
+ "model.language_model.layers.0.linear_attn.out_proj.weight": "model-00002-of-00002.safetensors",
196
+ "model.language_model.layers.1.linear_attn.in_proj_z.weight": "model-00002-of-00002.safetensors",
197
+ "model.language_model.layers.1.linear_attn.out_proj.weight": "model-00002-of-00002.safetensors",
198
+ "model.visual.merger.linear_fc2.weight": "model-00002-of-00002.safetensors",
199
+ "model.visual.blocks.0.mlp.linear_fc1.weight": "model-00002-of-00002.safetensors",
200
+ "model.visual.blocks.0.mlp.linear_fc2.weight": "model-00002-of-00002.safetensors",
201
+ "model.visual.blocks.1.mlp.linear_fc1.weight": "model-00002-of-00002.safetensors",
202
+ "model.visual.blocks.1.mlp.linear_fc2.weight": "model-00002-of-00002.safetensors",
203
+ "model.visual.blocks.10.mlp.linear_fc1.weight": "model-00002-of-00002.safetensors",
204
+ "model.visual.blocks.10.mlp.linear_fc2.weight": "model-00002-of-00002.safetensors",
205
+ "model.visual.blocks.11.mlp.linear_fc1.weight": "model-00002-of-00002.safetensors",
206
+ "model.visual.blocks.11.mlp.linear_fc2.weight": "model-00002-of-00002.safetensors",
207
+ "model.visual.blocks.12.mlp.linear_fc1.weight": "model-00002-of-00002.safetensors",
208
+ "model.visual.blocks.12.mlp.linear_fc2.weight": "model-00002-of-00002.safetensors",
209
+ "model.visual.blocks.13.mlp.linear_fc1.weight": "model-00002-of-00002.safetensors",
210
+ "model.visual.blocks.13.mlp.linear_fc2.weight": "model-00002-of-00002.safetensors",
211
+ "model.visual.blocks.14.mlp.linear_fc1.weight": "model-00002-of-00002.safetensors",
212
+ "model.visual.blocks.14.mlp.linear_fc2.weight": "model-00002-of-00002.safetensors",
213
+ "model.visual.blocks.15.mlp.linear_fc1.weight": "model-00002-of-00002.safetensors",
214
+ "model.visual.blocks.15.mlp.linear_fc2.weight": "model-00002-of-00002.safetensors",
215
+ "model.visual.blocks.16.mlp.linear_fc1.weight": "model-00002-of-00002.safetensors",
216
+ "model.visual.blocks.16.mlp.linear_fc2.weight": "model-00002-of-00002.safetensors",
217
+ "model.visual.blocks.17.mlp.linear_fc1.weight": "model-00002-of-00002.safetensors",
218
+ "model.visual.blocks.17.mlp.linear_fc2.weight": "model-00002-of-00002.safetensors",
219
+ "model.visual.blocks.18.mlp.linear_fc1.weight": "model-00002-of-00002.safetensors",
220
+ "model.visual.blocks.18.mlp.linear_fc2.weight": "model-00002-of-00002.safetensors",
221
+ "model.visual.blocks.19.mlp.linear_fc1.weight": "model-00002-of-00002.safetensors",
222
+ "model.visual.blocks.19.mlp.linear_fc2.weight": "model-00002-of-00002.safetensors",
223
+ "model.visual.blocks.2.mlp.linear_fc1.weight": "model-00002-of-00002.safetensors",
224
+ "model.visual.blocks.2.mlp.linear_fc2.weight": "model-00002-of-00002.safetensors",
225
+ "model.visual.blocks.20.mlp.linear_fc1.weight": "model-00002-of-00002.safetensors",
226
+ "model.visual.blocks.20.mlp.linear_fc2.weight": "model-00002-of-00002.safetensors",
227
+ "model.visual.blocks.21.mlp.linear_fc1.weight": "model-00002-of-00002.safetensors",
228
+ "model.visual.blocks.21.mlp.linear_fc2.weight": "model-00002-of-00002.safetensors",
229
+ "model.visual.blocks.22.mlp.linear_fc1.weight": "model-00002-of-00002.safetensors",
230
+ "model.visual.blocks.22.mlp.linear_fc2.weight": "model-00002-of-00002.safetensors",
231
+ "model.visual.blocks.23.mlp.linear_fc1.weight": "model-00002-of-00002.safetensors",
232
+ "model.visual.blocks.23.mlp.linear_fc2.weight": "model-00002-of-00002.safetensors",
233
+ "model.visual.blocks.3.mlp.linear_fc1.weight": "model-00002-of-00002.safetensors",
234
+ "model.visual.blocks.3.mlp.linear_fc2.weight": "model-00002-of-00002.safetensors",
235
+ "model.visual.blocks.4.mlp.linear_fc1.weight": "model-00002-of-00002.safetensors",
236
+ "model.visual.blocks.4.mlp.linear_fc2.weight": "model-00002-of-00002.safetensors",
237
+ "model.visual.blocks.5.mlp.linear_fc1.weight": "model-00002-of-00002.safetensors",
238
+ "model.visual.blocks.5.mlp.linear_fc2.weight": "model-00002-of-00002.safetensors",
239
+ "model.visual.blocks.6.mlp.linear_fc1.weight": "model-00002-of-00002.safetensors",
240
+ "model.visual.blocks.6.mlp.linear_fc2.weight": "model-00002-of-00002.safetensors",
241
+ "model.visual.blocks.7.mlp.linear_fc1.weight": "model-00002-of-00002.safetensors",
242
+ "model.visual.blocks.7.mlp.linear_fc2.weight": "model-00002-of-00002.safetensors",
243
+ "model.visual.blocks.8.mlp.linear_fc1.weight": "model-00002-of-00002.safetensors",
244
+ "model.visual.blocks.8.mlp.linear_fc2.weight": "model-00002-of-00002.safetensors",
245
+ "model.visual.blocks.9.mlp.linear_fc1.weight": "model-00002-of-00002.safetensors",
246
+ "model.visual.blocks.9.mlp.linear_fc2.weight": "model-00002-of-00002.safetensors",
247
+ "model.visual.blocks.0.attn.qkv.weight": "model-00002-of-00002.safetensors",
248
+ "model.visual.blocks.1.attn.qkv.weight": "model-00002-of-00002.safetensors",
249
+ "model.visual.blocks.10.attn.qkv.weight": "model-00002-of-00002.safetensors",
250
+ "model.visual.blocks.11.attn.qkv.weight": "model-00002-of-00002.safetensors",
251
+ "model.visual.blocks.12.attn.qkv.weight": "model-00002-of-00002.safetensors",
252
+ "model.visual.blocks.13.attn.qkv.weight": "model-00002-of-00002.safetensors",
253
+ "model.visual.blocks.14.attn.qkv.weight": "model-00002-of-00002.safetensors",
254
+ "model.visual.blocks.15.attn.qkv.weight": "model-00002-of-00002.safetensors",
255
+ "model.visual.blocks.16.attn.qkv.weight": "model-00002-of-00002.safetensors",
256
+ "model.visual.blocks.17.attn.qkv.weight": "model-00002-of-00002.safetensors",
257
+ "model.visual.blocks.18.attn.qkv.weight": "model-00002-of-00002.safetensors",
258
+ "model.visual.blocks.19.attn.qkv.weight": "model-00002-of-00002.safetensors",
259
+ "model.visual.blocks.2.attn.qkv.weight": "model-00002-of-00002.safetensors",
260
+ "model.visual.blocks.20.attn.qkv.weight": "model-00002-of-00002.safetensors",
261
+ "model.visual.blocks.21.attn.qkv.weight": "model-00002-of-00002.safetensors",
262
+ "model.visual.blocks.22.attn.qkv.weight": "model-00002-of-00002.safetensors",
263
+ "model.visual.blocks.23.attn.qkv.weight": "model-00002-of-00002.safetensors",
264
+ "model.visual.blocks.3.attn.qkv.weight": "model-00002-of-00002.safetensors",
265
+ "model.visual.blocks.4.attn.qkv.weight": "model-00002-of-00002.safetensors",
266
+ "model.visual.blocks.5.attn.qkv.weight": "model-00002-of-00002.safetensors",
267
+ "model.visual.blocks.6.attn.qkv.weight": "model-00002-of-00002.safetensors",
268
+ "model.visual.blocks.7.attn.qkv.weight": "model-00002-of-00002.safetensors",
269
+ "model.visual.blocks.8.attn.qkv.weight": "model-00002-of-00002.safetensors",
270
+ "model.visual.blocks.9.attn.qkv.weight": "model-00002-of-00002.safetensors",
271
+ "model.language_model.layers.11.self_attn.k_proj.weight": "model-00002-of-00002.safetensors",
272
+ "model.language_model.layers.11.self_attn.v_proj.weight": "model-00002-of-00002.safetensors",
273
+ "model.language_model.layers.19.self_attn.k_proj.weight": "model-00002-of-00002.safetensors",
274
+ "model.language_model.layers.19.self_attn.v_proj.weight": "model-00002-of-00002.safetensors",
275
+ "model.language_model.layers.7.self_attn.k_proj.weight": "model-00002-of-00002.safetensors",
276
+ "model.language_model.layers.7.self_attn.v_proj.weight": "model-00002-of-00002.safetensors",
277
+ "model.language_model.layers.23.self_attn.k_proj.weight": "model-00002-of-00002.safetensors",
278
+ "model.language_model.layers.23.self_attn.v_proj.weight": "model-00002-of-00002.safetensors",
279
+ "model.language_model.layers.31.self_attn.k_proj.weight": "model-00002-of-00002.safetensors",
280
+ "model.language_model.layers.31.self_attn.v_proj.weight": "model-00002-of-00002.safetensors",
281
+ "model.language_model.layers.15.self_attn.k_proj.weight": "model-00002-of-00002.safetensors",
282
+ "model.language_model.layers.15.self_attn.v_proj.weight": "model-00002-of-00002.safetensors",
283
+ "mtp.layers.0.self_attn.k_proj.weight": "model-00002-of-00002.safetensors",
284
+ "mtp.layers.0.self_attn.v_proj.weight": "model-00002-of-00002.safetensors",
285
+ "model.language_model.layers.3.self_attn.k_proj.weight": "model-00002-of-00002.safetensors",
286
+ "model.language_model.layers.3.self_attn.v_proj.weight": "model-00002-of-00002.safetensors",
287
+ "model.language_model.layers.27.self_attn.k_proj.weight": "model-00002-of-00002.safetensors",
288
+ "model.language_model.layers.27.self_attn.v_proj.weight": "model-00002-of-00002.safetensors",
289
+ "model.visual.pos_embed.weight": "model-00002-of-00002.safetensors",
290
+ "model.visual.patch_embed.proj.weight": "model-00002-of-00002.safetensors",
291
+ "model.visual.blocks.0.attn.proj.weight": "model-00002-of-00002.safetensors",
292
+ "model.visual.blocks.1.attn.proj.weight": "model-00002-of-00002.safetensors",
293
+ "model.visual.blocks.10.attn.proj.weight": "model-00002-of-00002.safetensors",
294
+ "model.visual.blocks.11.attn.proj.weight": "model-00002-of-00002.safetensors",
295
+ "model.visual.blocks.12.attn.proj.weight": "model-00002-of-00002.safetensors",
296
+ "model.visual.blocks.13.attn.proj.weight": "model-00002-of-00002.safetensors",
297
+ "model.visual.blocks.14.attn.proj.weight": "model-00002-of-00002.safetensors",
298
+ "model.visual.blocks.15.attn.proj.weight": "model-00002-of-00002.safetensors",
299
+ "model.visual.blocks.16.attn.proj.weight": "model-00002-of-00002.safetensors",
300
+ "model.visual.blocks.17.attn.proj.weight": "model-00002-of-00002.safetensors",
301
+ "model.visual.blocks.18.attn.proj.weight": "model-00002-of-00002.safetensors",
302
+ "model.visual.blocks.19.attn.proj.weight": "model-00002-of-00002.safetensors",
303
+ "model.visual.blocks.2.attn.proj.weight": "model-00002-of-00002.safetensors",
304
+ "model.visual.blocks.20.attn.proj.weight": "model-00002-of-00002.safetensors",
305
+ "model.visual.blocks.21.attn.proj.weight": "model-00002-of-00002.safetensors",
306
+ "model.visual.blocks.22.attn.proj.weight": "model-00002-of-00002.safetensors",
307
+ "model.visual.blocks.23.attn.proj.weight": "model-00002-of-00002.safetensors",
308
+ "model.visual.blocks.3.attn.proj.weight": "model-00002-of-00002.safetensors",
309
+ "model.visual.blocks.4.attn.proj.weight": "model-00002-of-00002.safetensors",
310
+ "model.visual.blocks.5.attn.proj.weight": "model-00002-of-00002.safetensors",
311
+ "model.visual.blocks.6.attn.proj.weight": "model-00002-of-00002.safetensors",
312
+ "model.visual.blocks.7.attn.proj.weight": "model-00002-of-00002.safetensors",
313
+ "model.visual.blocks.8.attn.proj.weight": "model-00002-of-00002.safetensors",
314
+ "model.visual.blocks.9.attn.proj.weight": "model-00002-of-00002.safetensors",
315
+ "model.language_model.layers.10.linear_attn.in_proj_b.weight": "model-00002-of-00002.safetensors",
316
+ "model.language_model.layers.10.linear_attn.in_proj_a.weight": "model-00002-of-00002.safetensors",
317
+ "model.language_model.layers.18.linear_attn.in_proj_b.weight": "model-00002-of-00002.safetensors",
318
+ "model.language_model.layers.18.linear_attn.in_proj_a.weight": "model-00002-of-00002.safetensors",
319
+ "model.language_model.layers.2.linear_attn.in_proj_b.weight": "model-00002-of-00002.safetensors",
320
+ "model.language_model.layers.2.linear_attn.in_proj_a.weight": "model-00002-of-00002.safetensors",
321
+ "model.language_model.layers.12.linear_attn.in_proj_b.weight": "model-00002-of-00002.safetensors",
322
+ "model.language_model.layers.12.linear_attn.in_proj_a.weight": "model-00002-of-00002.safetensors",
323
+ "model.language_model.layers.13.linear_attn.in_proj_b.weight": "model-00002-of-00002.safetensors",
324
+ "model.language_model.layers.13.linear_attn.in_proj_a.weight": "model-00002-of-00002.safetensors",
325
+ "model.language_model.layers.8.linear_attn.in_proj_b.weight": "model-00002-of-00002.safetensors",
326
+ "model.language_model.layers.8.linear_attn.in_proj_a.weight": "model-00002-of-00002.safetensors",
327
+ "model.language_model.layers.22.linear_attn.in_proj_b.weight": "model-00002-of-00002.safetensors",
328
+ "model.language_model.layers.22.linear_attn.in_proj_a.weight": "model-00002-of-00002.safetensors",
329
+ "model.language_model.layers.4.linear_attn.in_proj_b.weight": "model-00002-of-00002.safetensors",
330
+ "model.language_model.layers.4.linear_attn.in_proj_a.weight": "model-00002-of-00002.safetensors",
331
+ "model.language_model.layers.14.linear_attn.in_proj_b.weight": "model-00002-of-00002.safetensors",
332
+ "model.language_model.layers.14.linear_attn.in_proj_a.weight": "model-00002-of-00002.safetensors",
333
+ "model.language_model.layers.9.linear_attn.in_proj_b.weight": "model-00002-of-00002.safetensors",
334
+ "model.language_model.layers.9.linear_attn.in_proj_a.weight": "model-00002-of-00002.safetensors",
335
+ "model.language_model.layers.30.linear_attn.in_proj_b.weight": "model-00002-of-00002.safetensors",
336
+ "model.language_model.layers.30.linear_attn.in_proj_a.weight": "model-00002-of-00002.safetensors",
337
+ "model.language_model.layers.16.linear_attn.in_proj_b.weight": "model-00002-of-00002.safetensors",
338
+ "model.language_model.layers.16.linear_attn.in_proj_a.weight": "model-00002-of-00002.safetensors",
339
+ "model.language_model.layers.17.linear_attn.in_proj_b.weight": "model-00002-of-00002.safetensors",
340
+ "model.language_model.layers.17.linear_attn.in_proj_a.weight": "model-00002-of-00002.safetensors",
341
+ "model.language_model.layers.24.linear_attn.in_proj_b.weight": "model-00002-of-00002.safetensors",
342
+ "model.language_model.layers.24.linear_attn.in_proj_a.weight": "model-00002-of-00002.safetensors",
343
+ "model.language_model.layers.25.linear_attn.in_proj_b.weight": "model-00002-of-00002.safetensors",
344
+ "model.language_model.layers.25.linear_attn.in_proj_a.weight": "model-00002-of-00002.safetensors",
345
+ "model.language_model.layers.5.linear_attn.in_proj_b.weight": "model-00002-of-00002.safetensors",
346
+ "model.language_model.layers.5.linear_attn.in_proj_a.weight": "model-00002-of-00002.safetensors",
347
+ "model.language_model.layers.6.linear_attn.in_proj_b.weight": "model-00002-of-00002.safetensors",
348
+ "model.language_model.layers.6.linear_attn.in_proj_a.weight": "model-00002-of-00002.safetensors",
349
+ "model.language_model.layers.20.linear_attn.in_proj_b.weight": "model-00002-of-00002.safetensors",
350
+ "model.language_model.layers.20.linear_attn.in_proj_a.weight": "model-00002-of-00002.safetensors",
351
+ "model.language_model.layers.21.linear_attn.in_proj_b.weight": "model-00002-of-00002.safetensors",
352
+ "model.language_model.layers.21.linear_attn.in_proj_a.weight": "model-00002-of-00002.safetensors",
353
+ "model.language_model.layers.26.linear_attn.in_proj_b.weight": "model-00002-of-00002.safetensors",
354
+ "model.language_model.layers.26.linear_attn.in_proj_a.weight": "model-00002-of-00002.safetensors",
355
+ "model.language_model.layers.28.linear_attn.in_proj_b.weight": "model-00002-of-00002.safetensors",
356
+ "model.language_model.layers.28.linear_attn.in_proj_a.weight": "model-00002-of-00002.safetensors",
357
+ "model.language_model.layers.29.linear_attn.in_proj_b.weight": "model-00002-of-00002.safetensors",
358
+ "model.language_model.layers.29.linear_attn.in_proj_a.weight": "model-00002-of-00002.safetensors",
359
+ "model.language_model.layers.0.linear_attn.in_proj_b.weight": "model-00002-of-00002.safetensors",
360
+ "model.language_model.layers.0.linear_attn.in_proj_a.weight": "model-00002-of-00002.safetensors",
361
+ "model.language_model.layers.1.linear_attn.in_proj_b.weight": "model-00002-of-00002.safetensors",
362
+ "model.language_model.layers.1.linear_attn.in_proj_a.weight": "model-00002-of-00002.safetensors",
363
+ "model.language_model.layers.10.linear_attn.conv1d.weight": "model-00002-of-00002.safetensors",
364
+ "model.language_model.layers.2.linear_attn.conv1d.weight": "model-00002-of-00002.safetensors",
365
+ "model.language_model.layers.12.linear_attn.conv1d.weight": "model-00002-of-00002.safetensors",
366
+ "model.language_model.layers.13.linear_attn.conv1d.weight": "model-00002-of-00002.safetensors",
367
+ "model.language_model.layers.8.linear_attn.conv1d.weight": "model-00002-of-00002.safetensors",
368
+ "model.language_model.layers.9.linear_attn.conv1d.weight": "model-00002-of-00002.safetensors",
369
+ "model.language_model.layers.22.linear_attn.conv1d.weight": "model-00002-of-00002.safetensors",
370
+ "model.language_model.layers.4.linear_attn.conv1d.weight": "model-00002-of-00002.safetensors",
371
+ "model.language_model.layers.5.linear_attn.conv1d.weight": "model-00002-of-00002.safetensors",
372
+ "model.language_model.layers.14.linear_attn.conv1d.weight": "model-00002-of-00002.safetensors",
373
+ "model.language_model.layers.16.linear_attn.conv1d.weight": "model-00002-of-00002.safetensors",
374
+ "model.language_model.layers.30.linear_attn.conv1d.weight": "model-00002-of-00002.safetensors",
375
+ "model.language_model.layers.17.linear_attn.conv1d.weight": "model-00002-of-00002.safetensors",
376
+ "model.language_model.layers.18.linear_attn.conv1d.weight": "model-00002-of-00002.safetensors",
377
+ "model.language_model.layers.24.linear_attn.conv1d.weight": "model-00002-of-00002.safetensors",
378
+ "model.language_model.layers.25.linear_attn.conv1d.weight": "model-00002-of-00002.safetensors",
379
+ "model.language_model.layers.6.linear_attn.conv1d.weight": "model-00002-of-00002.safetensors",
380
+ "model.language_model.layers.20.linear_attn.conv1d.weight": "model-00002-of-00002.safetensors",
381
+ "model.language_model.layers.21.linear_attn.conv1d.weight": "model-00002-of-00002.safetensors",
382
+ "model.language_model.layers.26.linear_attn.conv1d.weight": "model-00002-of-00002.safetensors",
383
+ "model.language_model.layers.28.linear_attn.conv1d.weight": "model-00002-of-00002.safetensors",
384
+ "model.language_model.layers.29.linear_attn.conv1d.weight": "model-00002-of-00002.safetensors",
385
+ "model.language_model.layers.0.linear_attn.conv1d.weight": "model-00002-of-00002.safetensors",
386
+ "model.language_model.layers.1.linear_attn.conv1d.weight": "model-00002-of-00002.safetensors",
387
+ "model.visual.blocks.0.mlp.linear_fc1.bias": "model-00002-of-00002.safetensors",
388
+ "model.visual.blocks.1.mlp.linear_fc1.bias": "model-00002-of-00002.safetensors",
389
+ "model.visual.blocks.10.mlp.linear_fc1.bias": "model-00002-of-00002.safetensors",
390
+ "model.visual.blocks.11.mlp.linear_fc1.bias": "model-00002-of-00002.safetensors",
391
+ "model.visual.blocks.12.mlp.linear_fc1.bias": "model-00002-of-00002.safetensors",
392
+ "model.visual.blocks.13.mlp.linear_fc1.bias": "model-00002-of-00002.safetensors",
393
+ "model.visual.blocks.14.mlp.linear_fc1.bias": "model-00002-of-00002.safetensors",
394
+ "model.visual.blocks.15.mlp.linear_fc1.bias": "model-00002-of-00002.safetensors",
395
+ "model.visual.blocks.16.mlp.linear_fc1.bias": "model-00002-of-00002.safetensors",
396
+ "model.visual.blocks.17.mlp.linear_fc1.bias": "model-00002-of-00002.safetensors",
397
+ "model.visual.blocks.18.mlp.linear_fc1.bias": "model-00002-of-00002.safetensors",
398
+ "model.visual.blocks.19.mlp.linear_fc1.bias": "model-00002-of-00002.safetensors",
399
+ "model.visual.blocks.2.mlp.linear_fc1.bias": "model-00002-of-00002.safetensors",
400
+ "model.visual.blocks.20.mlp.linear_fc1.bias": "model-00002-of-00002.safetensors",
401
+ "model.visual.blocks.21.mlp.linear_fc1.bias": "model-00002-of-00002.safetensors",
402
+ "model.visual.blocks.22.mlp.linear_fc1.bias": "model-00002-of-00002.safetensors",
403
+ "model.visual.blocks.23.mlp.linear_fc1.bias": "model-00002-of-00002.safetensors",
404
+ "model.visual.blocks.3.mlp.linear_fc1.bias": "model-00002-of-00002.safetensors",
405
+ "model.visual.blocks.4.mlp.linear_fc1.bias": "model-00002-of-00002.safetensors",
406
+ "model.visual.blocks.5.mlp.linear_fc1.bias": "model-00002-of-00002.safetensors",
407
+ "model.visual.blocks.6.mlp.linear_fc1.bias": "model-00002-of-00002.safetensors",
408
+ "model.visual.blocks.7.mlp.linear_fc1.bias": "model-00002-of-00002.safetensors",
409
+ "model.visual.blocks.8.mlp.linear_fc1.bias": "model-00002-of-00002.safetensors",
410
+ "model.visual.blocks.9.mlp.linear_fc1.bias": "model-00002-of-00002.safetensors",
411
+ "model.visual.merger.linear_fc1.bias": "model-00002-of-00002.safetensors",
412
+ "model.visual.blocks.0.attn.qkv.bias": "model-00002-of-00002.safetensors",
413
+ "model.visual.blocks.1.attn.qkv.bias": "model-00002-of-00002.safetensors",
414
+ "model.visual.blocks.10.attn.qkv.bias": "model-00002-of-00002.safetensors",
415
+ "model.visual.blocks.11.attn.qkv.bias": "model-00002-of-00002.safetensors",
416
+ "model.visual.blocks.12.attn.qkv.bias": "model-00002-of-00002.safetensors",
417
+ "model.visual.blocks.13.attn.qkv.bias": "model-00002-of-00002.safetensors",
418
+ "model.visual.blocks.14.attn.qkv.bias": "model-00002-of-00002.safetensors",
419
+ "model.visual.blocks.15.attn.qkv.bias": "model-00002-of-00002.safetensors",
420
+ "model.visual.blocks.16.attn.qkv.bias": "model-00002-of-00002.safetensors",
421
+ "model.visual.blocks.17.attn.qkv.bias": "model-00002-of-00002.safetensors",
422
+ "model.visual.blocks.18.attn.qkv.bias": "model-00002-of-00002.safetensors",
423
+ "model.visual.blocks.19.attn.qkv.bias": "model-00002-of-00002.safetensors",
424
+ "model.visual.blocks.2.attn.qkv.bias": "model-00002-of-00002.safetensors",
425
+ "model.visual.blocks.20.attn.qkv.bias": "model-00002-of-00002.safetensors",
426
+ "model.visual.blocks.21.attn.qkv.bias": "model-00002-of-00002.safetensors",
427
+ "model.visual.blocks.22.attn.qkv.bias": "model-00002-of-00002.safetensors",
428
+ "model.visual.blocks.23.attn.qkv.bias": "model-00002-of-00002.safetensors",
429
+ "model.visual.blocks.3.attn.qkv.bias": "model-00002-of-00002.safetensors",
430
+ "model.visual.blocks.4.attn.qkv.bias": "model-00002-of-00002.safetensors",
431
+ "model.visual.blocks.5.attn.qkv.bias": "model-00002-of-00002.safetensors",
432
+ "model.visual.blocks.6.attn.qkv.bias": "model-00002-of-00002.safetensors",
433
+ "model.visual.blocks.7.attn.qkv.bias": "model-00002-of-00002.safetensors",
434
+ "model.visual.blocks.8.attn.qkv.bias": "model-00002-of-00002.safetensors",
435
+ "model.visual.blocks.9.attn.qkv.bias": "model-00002-of-00002.safetensors",
436
+ "model.language_model.layers.10.input_layernorm.weight": "model-00002-of-00002.safetensors",
437
+ "model.language_model.layers.10.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
438
+ "model.language_model.layers.11.input_layernorm.weight": "model-00002-of-00002.safetensors",
439
+ "model.language_model.layers.11.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
440
+ "model.language_model.layers.18.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
441
+ "model.language_model.layers.19.input_layernorm.weight": "model-00002-of-00002.safetensors",
442
+ "model.language_model.layers.19.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
443
+ "model.language_model.layers.2.input_layernorm.weight": "model-00002-of-00002.safetensors",
444
+ "model.language_model.layers.12.input_layernorm.weight": "model-00002-of-00002.safetensors",
445
+ "model.language_model.layers.12.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
446
+ "model.language_model.layers.13.input_layernorm.weight": "model-00002-of-00002.safetensors",
447
+ "model.language_model.layers.13.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
448
+ "model.language_model.layers.7.input_layernorm.weight": "model-00002-of-00002.safetensors",
449
+ "model.language_model.layers.8.input_layernorm.weight": "model-00002-of-00002.safetensors",
450
+ "model.language_model.layers.8.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
451
+ "model.language_model.layers.9.input_layernorm.weight": "model-00002-of-00002.safetensors",
452
+ "model.language_model.layers.21.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
453
+ "model.language_model.layers.22.input_layernorm.weight": "model-00002-of-00002.safetensors",
454
+ "model.language_model.layers.22.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
455
+ "model.language_model.layers.23.input_layernorm.weight": "model-00002-of-00002.safetensors",
456
+ "model.language_model.layers.23.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
457
+ "model.language_model.layers.31.input_layernorm.weight": "model-00002-of-00002.safetensors",
458
+ "model.language_model.layers.4.input_layernorm.weight": "model-00002-of-00002.safetensors",
459
+ "model.language_model.layers.4.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
460
+ "model.language_model.layers.14.input_layernorm.weight": "model-00002-of-00002.safetensors",
461
+ "model.language_model.layers.14.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
462
+ "model.language_model.layers.15.input_layernorm.weight": "model-00002-of-00002.safetensors",
463
+ "model.language_model.layers.15.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
464
+ "model.language_model.layers.16.input_layernorm.weight": "model-00002-of-00002.safetensors",
465
+ "model.language_model.layers.9.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
466
+ "mtp.layers.0.input_layernorm.weight": "model-00002-of-00002.safetensors",
467
+ "mtp.layers.0.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
468
+ "mtp.norm.weight": "model-00002-of-00002.safetensors",
469
+ "mtp.pre_fc_norm_embedding.weight": "model-00002-of-00002.safetensors",
470
+ "mtp.pre_fc_norm_hidden.weight": "model-00002-of-00002.safetensors",
471
+ "model.language_model.layers.3.input_layernorm.weight": "model-00002-of-00002.safetensors",
472
+ "model.language_model.layers.3.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
473
+ "model.language_model.layers.30.input_layernorm.weight": "model-00002-of-00002.safetensors",
474
+ "model.language_model.layers.30.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
475
+ "model.language_model.layers.31.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
476
+ "model.language_model.layers.16.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
477
+ "model.language_model.layers.17.input_layernorm.weight": "model-00002-of-00002.safetensors",
478
+ "model.language_model.layers.17.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
479
+ "model.language_model.layers.18.input_layernorm.weight": "model-00002-of-00002.safetensors",
480
+ "model.language_model.layers.24.input_layernorm.weight": "model-00002-of-00002.safetensors",
481
+ "model.language_model.layers.24.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
482
+ "model.language_model.layers.25.input_layernorm.weight": "model-00002-of-00002.safetensors",
483
+ "model.language_model.layers.25.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
484
+ "model.language_model.layers.5.input_layernorm.weight": "model-00002-of-00002.safetensors",
485
+ "model.language_model.layers.5.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
486
+ "model.language_model.layers.6.input_layernorm.weight": "model-00002-of-00002.safetensors",
487
+ "model.language_model.layers.6.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
488
+ "model.language_model.layers.7.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
489
+ "model.language_model.layers.2.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
490
+ "model.language_model.layers.20.input_layernorm.weight": "model-00002-of-00002.safetensors",
491
+ "model.language_model.layers.20.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
492
+ "model.language_model.layers.21.input_layernorm.weight": "model-00002-of-00002.safetensors",
493
+ "model.language_model.layers.26.input_layernorm.weight": "model-00002-of-00002.safetensors",
494
+ "model.language_model.layers.26.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
495
+ "model.language_model.layers.27.input_layernorm.weight": "model-00002-of-00002.safetensors",
496
+ "model.language_model.layers.27.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
497
+ "model.language_model.layers.28.input_layernorm.weight": "model-00002-of-00002.safetensors",
498
+ "model.language_model.layers.28.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
499
+ "model.language_model.layers.29.input_layernorm.weight": "model-00002-of-00002.safetensors",
500
+ "model.language_model.layers.29.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
501
+ "model.language_model.layers.0.input_layernorm.weight": "model-00002-of-00002.safetensors",
502
+ "model.language_model.layers.0.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
503
+ "model.language_model.layers.1.input_layernorm.weight": "model-00002-of-00002.safetensors",
504
+ "model.language_model.layers.1.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
505
+ "model.language_model.norm.weight": "model-00002-of-00002.safetensors",
506
+ "model.visual.merger.linear_fc2.bias": "model-00002-of-00002.safetensors",
507
+ "model.visual.blocks.0.attn.proj.bias": "model-00002-of-00002.safetensors",
508
+ "model.visual.blocks.0.mlp.linear_fc2.bias": "model-00002-of-00002.safetensors",
509
+ "model.visual.blocks.0.norm1.bias": "model-00002-of-00002.safetensors",
510
+ "model.visual.blocks.0.norm1.weight": "model-00002-of-00002.safetensors",
511
+ "model.visual.blocks.0.norm2.bias": "model-00002-of-00002.safetensors",
512
+ "model.visual.blocks.0.norm2.weight": "model-00002-of-00002.safetensors",
513
+ "model.visual.blocks.1.attn.proj.bias": "model-00002-of-00002.safetensors",
514
+ "model.visual.blocks.1.mlp.linear_fc2.bias": "model-00002-of-00002.safetensors",
515
+ "model.visual.blocks.1.norm1.bias": "model-00002-of-00002.safetensors",
516
+ "model.visual.blocks.1.norm1.weight": "model-00002-of-00002.safetensors",
517
+ "model.visual.blocks.1.norm2.bias": "model-00002-of-00002.safetensors",
518
+ "model.visual.blocks.1.norm2.weight": "model-00002-of-00002.safetensors",
519
+ "model.visual.blocks.10.attn.proj.bias": "model-00002-of-00002.safetensors",
520
+ "model.visual.blocks.10.mlp.linear_fc2.bias": "model-00002-of-00002.safetensors",
521
+ "model.visual.blocks.10.norm1.bias": "model-00002-of-00002.safetensors",
522
+ "model.visual.blocks.10.norm1.weight": "model-00002-of-00002.safetensors",
523
+ "model.visual.blocks.10.norm2.bias": "model-00002-of-00002.safetensors",
524
+ "model.visual.blocks.10.norm2.weight": "model-00002-of-00002.safetensors",
525
+ "model.visual.blocks.11.attn.proj.bias": "model-00002-of-00002.safetensors",
526
+ "model.visual.blocks.11.mlp.linear_fc2.bias": "model-00002-of-00002.safetensors",
527
+ "model.visual.blocks.11.norm1.bias": "model-00002-of-00002.safetensors",
528
+ "model.visual.blocks.11.norm1.weight": "model-00002-of-00002.safetensors",
529
+ "model.visual.blocks.11.norm2.bias": "model-00002-of-00002.safetensors",
530
+ "model.visual.blocks.11.norm2.weight": "model-00002-of-00002.safetensors",
531
+ "model.visual.blocks.12.attn.proj.bias": "model-00002-of-00002.safetensors",
532
+ "model.visual.blocks.12.mlp.linear_fc2.bias": "model-00002-of-00002.safetensors",
533
+ "model.visual.blocks.12.norm1.bias": "model-00002-of-00002.safetensors",
534
+ "model.visual.blocks.12.norm1.weight": "model-00002-of-00002.safetensors",
535
+ "model.visual.blocks.12.norm2.bias": "model-00002-of-00002.safetensors",
536
+ "model.visual.blocks.12.norm2.weight": "model-00002-of-00002.safetensors",
537
+ "model.visual.blocks.13.attn.proj.bias": "model-00002-of-00002.safetensors",
538
+ "model.visual.blocks.13.mlp.linear_fc2.bias": "model-00002-of-00002.safetensors",
539
+ "model.visual.blocks.13.norm1.bias": "model-00002-of-00002.safetensors",
540
+ "model.visual.blocks.13.norm1.weight": "model-00002-of-00002.safetensors",
541
+ "model.visual.blocks.13.norm2.bias": "model-00002-of-00002.safetensors",
542
+ "model.visual.blocks.13.norm2.weight": "model-00002-of-00002.safetensors",
543
+ "model.visual.blocks.14.attn.proj.bias": "model-00002-of-00002.safetensors",
544
+ "model.visual.blocks.14.mlp.linear_fc2.bias": "model-00002-of-00002.safetensors",
545
+ "model.visual.blocks.14.norm1.bias": "model-00002-of-00002.safetensors",
546
+ "model.visual.blocks.14.norm1.weight": "model-00002-of-00002.safetensors",
547
+ "model.visual.blocks.14.norm2.bias": "model-00002-of-00002.safetensors",
548
+ "model.visual.blocks.14.norm2.weight": "model-00002-of-00002.safetensors",
549
+ "model.visual.blocks.15.attn.proj.bias": "model-00002-of-00002.safetensors",
550
+ "model.visual.blocks.15.mlp.linear_fc2.bias": "model-00002-of-00002.safetensors",
551
+ "model.visual.blocks.15.norm1.bias": "model-00002-of-00002.safetensors",
552
+ "model.visual.blocks.15.norm1.weight": "model-00002-of-00002.safetensors",
553
+ "model.visual.blocks.15.norm2.bias": "model-00002-of-00002.safetensors",
554
+ "model.visual.blocks.15.norm2.weight": "model-00002-of-00002.safetensors",
555
+ "model.visual.blocks.16.attn.proj.bias": "model-00002-of-00002.safetensors",
556
+ "model.visual.blocks.16.mlp.linear_fc2.bias": "model-00002-of-00002.safetensors",
557
+ "model.visual.blocks.16.norm1.bias": "model-00002-of-00002.safetensors",
558
+ "model.visual.blocks.16.norm1.weight": "model-00002-of-00002.safetensors",
559
+ "model.visual.blocks.16.norm2.bias": "model-00002-of-00002.safetensors",
560
+ "model.visual.blocks.16.norm2.weight": "model-00002-of-00002.safetensors",
561
+ "model.visual.blocks.17.attn.proj.bias": "model-00002-of-00002.safetensors",
562
+ "model.visual.blocks.17.mlp.linear_fc2.bias": "model-00002-of-00002.safetensors",
563
+ "model.visual.blocks.17.norm1.bias": "model-00002-of-00002.safetensors",
564
+ "model.visual.blocks.17.norm1.weight": "model-00002-of-00002.safetensors",
565
+ "model.visual.blocks.17.norm2.bias": "model-00002-of-00002.safetensors",
566
+ "model.visual.blocks.17.norm2.weight": "model-00002-of-00002.safetensors",
567
+ "model.visual.blocks.18.attn.proj.bias": "model-00002-of-00002.safetensors",
568
+ "model.visual.blocks.18.mlp.linear_fc2.bias": "model-00002-of-00002.safetensors",
569
+ "model.visual.blocks.18.norm1.bias": "model-00002-of-00002.safetensors",
570
+ "model.visual.blocks.18.norm1.weight": "model-00002-of-00002.safetensors",
571
+ "model.visual.blocks.18.norm2.bias": "model-00002-of-00002.safetensors",
572
+ "model.visual.blocks.18.norm2.weight": "model-00002-of-00002.safetensors",
573
+ "model.visual.blocks.19.attn.proj.bias": "model-00002-of-00002.safetensors",
574
+ "model.visual.blocks.19.mlp.linear_fc2.bias": "model-00002-of-00002.safetensors",
575
+ "model.visual.blocks.19.norm1.bias": "model-00002-of-00002.safetensors",
576
+ "model.visual.blocks.19.norm1.weight": "model-00002-of-00002.safetensors",
577
+ "model.visual.blocks.19.norm2.bias": "model-00002-of-00002.safetensors",
578
+ "model.visual.blocks.19.norm2.weight": "model-00002-of-00002.safetensors",
579
+ "model.visual.blocks.2.attn.proj.bias": "model-00002-of-00002.safetensors",
580
+ "model.visual.blocks.2.mlp.linear_fc2.bias": "model-00002-of-00002.safetensors",
581
+ "model.visual.blocks.2.norm1.bias": "model-00002-of-00002.safetensors",
582
+ "model.visual.blocks.2.norm1.weight": "model-00002-of-00002.safetensors",
583
+ "model.visual.blocks.2.norm2.bias": "model-00002-of-00002.safetensors",
584
+ "model.visual.blocks.2.norm2.weight": "model-00002-of-00002.safetensors",
585
+ "model.visual.blocks.20.attn.proj.bias": "model-00002-of-00002.safetensors",
586
+ "model.visual.blocks.20.mlp.linear_fc2.bias": "model-00002-of-00002.safetensors",
587
+ "model.visual.blocks.20.norm1.bias": "model-00002-of-00002.safetensors",
588
+ "model.visual.blocks.20.norm1.weight": "model-00002-of-00002.safetensors",
589
+ "model.visual.blocks.20.norm2.bias": "model-00002-of-00002.safetensors",
590
+ "model.visual.blocks.20.norm2.weight": "model-00002-of-00002.safetensors",
591
+ "model.visual.blocks.21.attn.proj.bias": "model-00002-of-00002.safetensors",
592
+ "model.visual.blocks.21.mlp.linear_fc2.bias": "model-00002-of-00002.safetensors",
593
+ "model.visual.blocks.21.norm1.bias": "model-00002-of-00002.safetensors",
594
+ "model.visual.blocks.21.norm1.weight": "model-00002-of-00002.safetensors",
595
+ "model.visual.blocks.21.norm2.bias": "model-00002-of-00002.safetensors",
596
+ "model.visual.blocks.21.norm2.weight": "model-00002-of-00002.safetensors",
597
+ "model.visual.blocks.22.attn.proj.bias": "model-00002-of-00002.safetensors",
598
+ "model.visual.blocks.22.mlp.linear_fc2.bias": "model-00002-of-00002.safetensors",
599
+ "model.visual.blocks.22.norm1.bias": "model-00002-of-00002.safetensors",
600
+ "model.visual.blocks.22.norm1.weight": "model-00002-of-00002.safetensors",
601
+ "model.visual.blocks.22.norm2.bias": "model-00002-of-00002.safetensors",
602
+ "model.visual.blocks.22.norm2.weight": "model-00002-of-00002.safetensors",
603
+ "model.visual.blocks.23.attn.proj.bias": "model-00002-of-00002.safetensors",
604
+ "model.visual.blocks.23.mlp.linear_fc2.bias": "model-00002-of-00002.safetensors",
605
+ "model.visual.blocks.23.norm1.bias": "model-00002-of-00002.safetensors",
606
+ "model.visual.blocks.23.norm1.weight": "model-00002-of-00002.safetensors",
607
+ "model.visual.blocks.23.norm2.bias": "model-00002-of-00002.safetensors",
608
+ "model.visual.blocks.23.norm2.weight": "model-00002-of-00002.safetensors",
609
+ "model.visual.blocks.3.attn.proj.bias": "model-00002-of-00002.safetensors",
610
+ "model.visual.blocks.3.mlp.linear_fc2.bias": "model-00002-of-00002.safetensors",
611
+ "model.visual.blocks.3.norm1.bias": "model-00002-of-00002.safetensors",
612
+ "model.visual.blocks.3.norm1.weight": "model-00002-of-00002.safetensors",
613
+ "model.visual.blocks.3.norm2.bias": "model-00002-of-00002.safetensors",
614
+ "model.visual.blocks.3.norm2.weight": "model-00002-of-00002.safetensors",
615
+ "model.visual.blocks.4.attn.proj.bias": "model-00002-of-00002.safetensors",
616
+ "model.visual.blocks.4.mlp.linear_fc2.bias": "model-00002-of-00002.safetensors",
617
+ "model.visual.blocks.4.norm1.bias": "model-00002-of-00002.safetensors",
618
+ "model.visual.blocks.4.norm1.weight": "model-00002-of-00002.safetensors",
619
+ "model.visual.blocks.4.norm2.bias": "model-00002-of-00002.safetensors",
620
+ "model.visual.blocks.4.norm2.weight": "model-00002-of-00002.safetensors",
621
+ "model.visual.blocks.5.attn.proj.bias": "model-00002-of-00002.safetensors",
622
+ "model.visual.blocks.5.mlp.linear_fc2.bias": "model-00002-of-00002.safetensors",
623
+ "model.visual.blocks.5.norm1.bias": "model-00002-of-00002.safetensors",
624
+ "model.visual.blocks.5.norm1.weight": "model-00002-of-00002.safetensors",
625
+ "model.visual.blocks.5.norm2.bias": "model-00002-of-00002.safetensors",
626
+ "model.visual.blocks.5.norm2.weight": "model-00002-of-00002.safetensors",
627
+ "model.visual.blocks.6.attn.proj.bias": "model-00002-of-00002.safetensors",
628
+ "model.visual.blocks.6.mlp.linear_fc2.bias": "model-00002-of-00002.safetensors",
629
+ "model.visual.blocks.6.norm1.bias": "model-00002-of-00002.safetensors",
630
+ "model.visual.blocks.6.norm1.weight": "model-00002-of-00002.safetensors",
631
+ "model.visual.blocks.6.norm2.bias": "model-00002-of-00002.safetensors",
632
+ "model.visual.blocks.6.norm2.weight": "model-00002-of-00002.safetensors",
633
+ "model.visual.blocks.7.attn.proj.bias": "model-00002-of-00002.safetensors",
634
+ "model.visual.blocks.7.mlp.linear_fc2.bias": "model-00002-of-00002.safetensors",
635
+ "model.visual.blocks.7.norm1.bias": "model-00002-of-00002.safetensors",
636
+ "model.visual.blocks.7.norm1.weight": "model-00002-of-00002.safetensors",
637
+ "model.visual.blocks.7.norm2.bias": "model-00002-of-00002.safetensors",
638
+ "model.visual.blocks.7.norm2.weight": "model-00002-of-00002.safetensors",
639
+ "model.visual.blocks.8.attn.proj.bias": "model-00002-of-00002.safetensors",
640
+ "model.visual.blocks.8.mlp.linear_fc2.bias": "model-00002-of-00002.safetensors",
641
+ "model.visual.blocks.8.norm1.bias": "model-00002-of-00002.safetensors",
642
+ "model.visual.blocks.8.norm1.weight": "model-00002-of-00002.safetensors",
643
+ "model.visual.blocks.8.norm2.bias": "model-00002-of-00002.safetensors",
644
+ "model.visual.blocks.8.norm2.weight": "model-00002-of-00002.safetensors",
645
+ "model.visual.blocks.9.attn.proj.bias": "model-00002-of-00002.safetensors",
646
+ "model.visual.blocks.9.mlp.linear_fc2.bias": "model-00002-of-00002.safetensors",
647
+ "model.visual.blocks.9.norm1.bias": "model-00002-of-00002.safetensors",
648
+ "model.visual.blocks.9.norm1.weight": "model-00002-of-00002.safetensors",
649
+ "model.visual.blocks.9.norm2.bias": "model-00002-of-00002.safetensors",
650
+ "model.visual.blocks.9.norm2.weight": "model-00002-of-00002.safetensors",
651
+ "model.visual.merger.norm.bias": "model-00002-of-00002.safetensors",
652
+ "model.visual.merger.norm.weight": "model-00002-of-00002.safetensors",
653
+ "model.visual.patch_embed.proj.bias": "model-00002-of-00002.safetensors",
654
+ "model.language_model.layers.10.linear_attn.norm.weight": "model-00002-of-00002.safetensors",
655
+ "model.language_model.layers.11.self_attn.k_norm.weight": "model-00002-of-00002.safetensors",
656
+ "model.language_model.layers.11.self_attn.q_norm.weight": "model-00002-of-00002.safetensors",
657
+ "model.language_model.layers.18.linear_attn.norm.weight": "model-00002-of-00002.safetensors",
658
+ "model.language_model.layers.19.self_attn.k_norm.weight": "model-00002-of-00002.safetensors",
659
+ "model.language_model.layers.19.self_attn.q_norm.weight": "model-00002-of-00002.safetensors",
660
+ "model.language_model.layers.2.linear_attn.norm.weight": "model-00002-of-00002.safetensors",
661
+ "model.language_model.layers.12.linear_attn.norm.weight": "model-00002-of-00002.safetensors",
662
+ "model.language_model.layers.13.linear_attn.norm.weight": "model-00002-of-00002.safetensors",
663
+ "model.language_model.layers.7.self_attn.k_norm.weight": "model-00002-of-00002.safetensors",
664
+ "model.language_model.layers.7.self_attn.q_norm.weight": "model-00002-of-00002.safetensors",
665
+ "model.language_model.layers.8.linear_attn.norm.weight": "model-00002-of-00002.safetensors",
666
+ "model.language_model.layers.22.linear_attn.norm.weight": "model-00002-of-00002.safetensors",
667
+ "model.language_model.layers.23.self_attn.k_norm.weight": "model-00002-of-00002.safetensors",
668
+ "model.language_model.layers.31.self_attn.k_norm.weight": "model-00002-of-00002.safetensors",
669
+ "model.language_model.layers.31.self_attn.q_norm.weight": "model-00002-of-00002.safetensors",
670
+ "model.language_model.layers.4.linear_attn.norm.weight": "model-00002-of-00002.safetensors",
671
+ "model.language_model.layers.14.linear_attn.norm.weight": "model-00002-of-00002.safetensors",
672
+ "model.language_model.layers.15.self_attn.k_norm.weight": "model-00002-of-00002.safetensors",
673
+ "model.language_model.layers.15.self_attn.q_norm.weight": "model-00002-of-00002.safetensors",
674
+ "model.language_model.layers.9.linear_attn.norm.weight": "model-00002-of-00002.safetensors",
675
+ "mtp.layers.0.self_attn.k_norm.weight": "model-00002-of-00002.safetensors",
676
+ "mtp.layers.0.self_attn.q_norm.weight": "model-00002-of-00002.safetensors",
677
+ "model.language_model.layers.3.self_attn.k_norm.weight": "model-00002-of-00002.safetensors",
678
+ "model.language_model.layers.3.self_attn.q_norm.weight": "model-00002-of-00002.safetensors",
679
+ "model.language_model.layers.30.linear_attn.norm.weight": "model-00002-of-00002.safetensors",
680
+ "model.language_model.layers.16.linear_attn.norm.weight": "model-00002-of-00002.safetensors",
681
+ "model.language_model.layers.17.linear_attn.norm.weight": "model-00002-of-00002.safetensors",
682
+ "model.language_model.layers.23.self_attn.q_norm.weight": "model-00002-of-00002.safetensors",
683
+ "model.language_model.layers.24.linear_attn.norm.weight": "model-00002-of-00002.safetensors",
684
+ "model.language_model.layers.25.linear_attn.norm.weight": "model-00002-of-00002.safetensors",
685
+ "model.language_model.layers.5.linear_attn.norm.weight": "model-00002-of-00002.safetensors",
686
+ "model.language_model.layers.6.linear_attn.norm.weight": "model-00002-of-00002.safetensors",
687
+ "model.language_model.layers.20.linear_attn.norm.weight": "model-00002-of-00002.safetensors",
688
+ "model.language_model.layers.21.linear_attn.norm.weight": "model-00002-of-00002.safetensors",
689
+ "model.language_model.layers.26.linear_attn.norm.weight": "model-00002-of-00002.safetensors",
690
+ "model.language_model.layers.27.self_attn.k_norm.weight": "model-00002-of-00002.safetensors",
691
+ "model.language_model.layers.27.self_attn.q_norm.weight": "model-00002-of-00002.safetensors",
692
+ "model.language_model.layers.28.linear_attn.norm.weight": "model-00002-of-00002.safetensors",
693
+ "model.language_model.layers.29.linear_attn.norm.weight": "model-00002-of-00002.safetensors",
694
+ "model.language_model.layers.0.linear_attn.norm.weight": "model-00002-of-00002.safetensors",
695
+ "model.language_model.layers.1.linear_attn.norm.weight": "model-00002-of-00002.safetensors",
696
+ "model.language_model.layers.10.linear_attn.A_log": "model-00002-of-00002.safetensors",
697
+ "model.language_model.layers.12.linear_attn.A_log": "model-00002-of-00002.safetensors",
698
+ "model.language_model.layers.2.linear_attn.A_log": "model-00002-of-00002.safetensors",
699
+ "model.language_model.layers.13.linear_attn.A_log": "model-00002-of-00002.safetensors",
700
+ "model.language_model.layers.14.linear_attn.A_log": "model-00002-of-00002.safetensors",
701
+ "model.language_model.layers.8.linear_attn.A_log": "model-00002-of-00002.safetensors",
702
+ "model.language_model.layers.9.linear_attn.A_log": "model-00002-of-00002.safetensors",
703
+ "model.language_model.layers.22.linear_attn.A_log": "model-00002-of-00002.safetensors",
704
+ "model.language_model.layers.4.linear_attn.A_log": "model-00002-of-00002.safetensors",
705
+ "model.language_model.layers.5.linear_attn.A_log": "model-00002-of-00002.safetensors",
706
+ "model.language_model.layers.16.linear_attn.A_log": "model-00002-of-00002.safetensors",
707
+ "model.language_model.layers.30.linear_attn.A_log": "model-00002-of-00002.safetensors",
708
+ "model.language_model.layers.17.linear_attn.A_log": "model-00002-of-00002.safetensors",
709
+ "model.language_model.layers.18.linear_attn.A_log": "model-00002-of-00002.safetensors",
710
+ "model.language_model.layers.24.linear_attn.A_log": "model-00002-of-00002.safetensors",
711
+ "model.language_model.layers.25.linear_attn.A_log": "model-00002-of-00002.safetensors",
712
+ "model.language_model.layers.6.linear_attn.A_log": "model-00002-of-00002.safetensors",
713
+ "model.language_model.layers.20.linear_attn.A_log": "model-00002-of-00002.safetensors",
714
+ "model.language_model.layers.21.linear_attn.A_log": "model-00002-of-00002.safetensors",
715
+ "model.language_model.layers.26.linear_attn.A_log": "model-00002-of-00002.safetensors",
716
+ "model.language_model.layers.28.linear_attn.A_log": "model-00002-of-00002.safetensors",
717
+ "model.language_model.layers.29.linear_attn.A_log": "model-00002-of-00002.safetensors",
718
+ "model.language_model.layers.0.linear_attn.A_log": "model-00002-of-00002.safetensors",
719
+ "model.language_model.layers.1.linear_attn.A_log": "model-00002-of-00002.safetensors",
720
+ "model.language_model.layers.10.linear_attn.dt_bias": "model-00002-of-00002.safetensors",
721
+ "model.language_model.layers.2.linear_attn.dt_bias": "model-00002-of-00002.safetensors",
722
+ "model.language_model.layers.12.linear_attn.dt_bias": "model-00002-of-00002.safetensors",
723
+ "model.language_model.layers.13.linear_attn.dt_bias": "model-00002-of-00002.safetensors",
724
+ "model.language_model.layers.8.linear_attn.dt_bias": "model-00002-of-00002.safetensors",
725
+ "model.language_model.layers.9.linear_attn.dt_bias": "model-00002-of-00002.safetensors",
726
+ "model.language_model.layers.22.linear_attn.dt_bias": "model-00002-of-00002.safetensors",
727
+ "model.language_model.layers.4.linear_attn.dt_bias": "model-00002-of-00002.safetensors",
728
+ "model.language_model.layers.5.linear_attn.dt_bias": "model-00002-of-00002.safetensors",
729
+ "model.language_model.layers.14.linear_attn.dt_bias": "model-00002-of-00002.safetensors",
730
+ "model.language_model.layers.16.linear_attn.dt_bias": "model-00002-of-00002.safetensors",
731
+ "model.language_model.layers.30.linear_attn.dt_bias": "model-00002-of-00002.safetensors",
732
+ "model.language_model.layers.17.linear_attn.dt_bias": "model-00002-of-00002.safetensors",
733
+ "model.language_model.layers.18.linear_attn.dt_bias": "model-00002-of-00002.safetensors",
734
+ "model.language_model.layers.24.linear_attn.dt_bias": "model-00002-of-00002.safetensors",
735
+ "model.language_model.layers.25.linear_attn.dt_bias": "model-00002-of-00002.safetensors",
736
+ "model.language_model.layers.6.linear_attn.dt_bias": "model-00002-of-00002.safetensors",
737
+ "model.language_model.layers.20.linear_attn.dt_bias": "model-00002-of-00002.safetensors",
738
+ "model.language_model.layers.21.linear_attn.dt_bias": "model-00002-of-00002.safetensors",
739
+ "model.language_model.layers.26.linear_attn.dt_bias": "model-00002-of-00002.safetensors",
740
+ "model.language_model.layers.28.linear_attn.dt_bias": "model-00002-of-00002.safetensors",
741
+ "model.language_model.layers.29.linear_attn.dt_bias": "model-00002-of-00002.safetensors",
742
+ "model.language_model.layers.0.linear_attn.dt_bias": "model-00002-of-00002.safetensors",
743
+ "model.language_model.layers.1.linear_attn.dt_bias": "model-00002-of-00002.safetensors"
744
+ }
745
+ }
processor_config.json ADDED
@@ -0,0 +1,63 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "image_processor": {
3
+ "data_format": "channels_first",
4
+ "do_convert_rgb": true,
5
+ "do_normalize": true,
6
+ "do_rescale": true,
7
+ "do_resize": true,
8
+ "image_mean": [
9
+ 0.5,
10
+ 0.5,
11
+ 0.5
12
+ ],
13
+ "image_processor_type": "Qwen2VLImageProcessorFast",
14
+ "image_std": [
15
+ 0.5,
16
+ 0.5,
17
+ 0.5
18
+ ],
19
+ "merge_size": 2,
20
+ "patch_size": 16,
21
+ "resample": 3,
22
+ "rescale_factor": 0.00392156862745098,
23
+ "size": {
24
+ "longest_edge": 16777216,
25
+ "shortest_edge": 65536
26
+ },
27
+ "temporal_patch_size": 2
28
+ },
29
+ "processor_class": "Qwen3VLProcessor",
30
+ "video_processor": {
31
+ "data_format": "channels_first",
32
+ "default_to_square": true,
33
+ "do_convert_rgb": true,
34
+ "do_normalize": true,
35
+ "do_rescale": true,
36
+ "do_resize": true,
37
+ "do_sample_frames": true,
38
+ "fps": 2,
39
+ "image_mean": [
40
+ 0.5,
41
+ 0.5,
42
+ 0.5
43
+ ],
44
+ "image_std": [
45
+ 0.5,
46
+ 0.5,
47
+ 0.5
48
+ ],
49
+ "max_frames": 768,
50
+ "merge_size": 2,
51
+ "min_frames": 4,
52
+ "patch_size": 16,
53
+ "resample": 3,
54
+ "rescale_factor": 0.00392156862745098,
55
+ "return_metadata": false,
56
+ "size": {
57
+ "longest_edge": 25165824,
58
+ "shortest_edge": 4096
59
+ },
60
+ "temporal_patch_size": 2,
61
+ "video_processor_type": "Qwen3VLVideoProcessor"
62
+ }
63
+ }
tokenizer.json ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:cb5c6475e730b6520efd128ddcd4788c4f0046621f19a2662ef97075d172b385
3
+ size 19989621
tokenizer_config.json ADDED
@@ -0,0 +1,300 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "add_prefix_space": false,
3
+ "audio_bos_token": "<|audio_start|>",
4
+ "audio_eos_token": "<|audio_end|>",
5
+ "audio_token": "<|audio_pad|>",
6
+ "backend": "tokenizers",
7
+ "bos_token": null,
8
+ "clean_up_tokenization_spaces": false,
9
+ "eos_token": "<|im_end|>",
10
+ "errors": "replace",
11
+ "image_token": "<|image_pad|>",
12
+ "is_local": false,
13
+ "model_max_length": 262144,
14
+ "model_specific_special_tokens": {
15
+ "audio_bos_token": "<|audio_start|>",
16
+ "audio_eos_token": "<|audio_end|>",
17
+ "audio_token": "<|audio_pad|>",
18
+ "image_token": "<|image_pad|>",
19
+ "video_token": "<|video_pad|>",
20
+ "vision_bos_token": "<|vision_start|>",
21
+ "vision_eos_token": "<|vision_end|>"
22
+ },
23
+ "pad_token": "<|vision_pad|>",
24
+ "padding_side": "right",
25
+ "pretokenize_regex": "(?i:'s|'t|'re|'ve|'m|'ll|'d)|[^\\r\\n\\p{L}\\p{N}]?[\\p{L}\\p{M}]+|\\p{N}| ?[^\\s\\p{L}\\p{M}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+",
26
+ "processor_class": "Qwen3VLProcessor",
27
+ "split_special_tokens": false,
28
+ "tokenizer_class": "TokenizersBackend",
29
+ "unk_token": null,
30
+ "video_token": "<|video_pad|>",
31
+ "vision_bos_token": "<|vision_start|>",
32
+ "vision_eos_token": "<|vision_end|>",
33
+ "chat_template": "{%- if tools %}\n {{- '<|im_start|>system\\n' }}\n {%- if messages[0].role == 'system' %}\n {{- messages[0].content + '\\n\\n' }}\n {%- endif %}\n {{- \"# Tools\\n\\nYou may call one or more functions to assist with the user query.\\n\\nYou are provided with function signatures within <tools></tools> XML tags:\\n<tools>\" }}\n {%- for tool in tools %}\n {{- \"\\n\" }}\n {{- tool | tojson }}\n {%- endfor %}\n {{- \"\\n</tools>\\n\\nFor each function call, return a json object with function name and arguments within <tool_call></tool_call> XML tags:\\n<tool_call>\\n{\\\"name\\\": <function-name>, \\\"arguments\\\": <args-json-object>}\\n</tool_call><|im_end|>\\n\" }}\n{%- else %}\n {%- if messages[0].role == 'system' %}\n {{- '<|im_start|>system\\n' + messages[0].content + '<|im_end|>\\n' }}\n {%- endif %}\n{%- endif %}\n{%- set ns = namespace(multi_step_tool=true, last_query_index=messages|length - 1) %}\n{%- for message in messages[::-1] %}\n {%- set index = (messages|length - 1) - loop.index0 %}\n {%- if ns.multi_step_tool and message.role == \"user\" and message.content is string and not(message.content.startswith('<tool_response>') and message.content.endswith('</tool_response>')) %}\n {%- set ns.multi_step_tool = false %}\n {%- set ns.last_query_index = index %}\n {%- endif %}\n{%- endfor %}\n{%- for message in messages %}\n {%- if message.content is string %}\n {%- set content = message.content %}\n {%- else %}\n {%- set content = '' %}\n {%- endif %}\n {%- if (message.role == \"user\") or (message.role == \"system\" and not loop.first) %}\n {{- '<|im_start|>' + message.role + '\\n' + content + '<|im_end|>' + '\\n' }}\n {%- elif message.role == \"assistant\" %}\n {%- set reasoning_content = '' %}\n {%- if message.reasoning_content is string %}\n {%- set reasoning_content = message.reasoning_content %}\n {%- else %}\n {%- if '</think>' in content %}\n {%- set reasoning_content = content.split('</think>')[0].rstrip('\\n').split('<think>')[-1].lstrip('\\n') %}\n {%- set content = content.split('</think>')[-1].lstrip('\\n') %}\n {%- endif %}\n {%- endif %}\n {%- if loop.index0 > ns.last_query_index %}\n {%- if loop.last or (not loop.last and reasoning_content) %}\n {{- '<|im_start|>' + message.role + '\\n<think>\\n' + reasoning_content.strip('\\n') + '\\n</think>\\n\\n' + content.lstrip('\\n') }}\n {%- else %}\n {{- '<|im_start|>' + message.role + '\\n' + content }}\n {%- endif %}\n {%- else %}\n {{- '<|im_start|>' + message.role + '\\n' + content }}\n {%- endif %}\n {%- if message.tool_calls %}\n {%- for tool_call in message.tool_calls %}\n {%- if (loop.first and content) or (not loop.first) %}\n {{- '\\n' }}\n {%- endif %}\n {%- if tool_call.function %}\n {%- set tool_call = tool_call.function %}\n {%- endif %}\n {{- '<tool_call>\\n{\"name\": \"' }}\n {{- tool_call.name }}\n {{- '\", \"arguments\": ' }}\n {%- if tool_call.arguments is string %}\n {{- tool_call.arguments }}\n {%- else %}\n {{- tool_call.arguments | tojson }}\n {%- endif %}\n {{- '}\\n</tool_call>' }}\n {%- endfor %}\n {%- endif %}\n {{- '<|im_end|>\\n' }}\n {%- elif message.role == \"tool\" %}\n {%- if loop.first or (messages[loop.index0 - 1].role != \"tool\") %}\n {{- '<|im_start|>user' }}\n {%- endif %}\n {{- '\\n<tool_response>\\n' }}\n {{- content }}\n {{- '\\n</tool_response>' }}\n {%- if loop.last or (messages[loop.index0 + 1].role != \"tool\") %}\n {{- '<|im_end|>\\n' }}\n {%- endif %}\n {%- endif %}\n{%- endfor %}\n{%- if add_generation_prompt %}\n {{- '<|im_start|>assistant\n<think>\n' }}\n{%- endif %}",
34
+ "added_tokens_decoder": {
35
+ "248044": {
36
+ "content": "<|endoftext|>",
37
+ "lstrip": false,
38
+ "normalized": false,
39
+ "rstrip": false,
40
+ "single_word": false,
41
+ "special": true
42
+ },
43
+ "248045": {
44
+ "content": "<|im_start|>",
45
+ "lstrip": false,
46
+ "normalized": false,
47
+ "rstrip": false,
48
+ "single_word": false,
49
+ "special": true
50
+ },
51
+ "248046": {
52
+ "content": "<|im_end|>",
53
+ "lstrip": false,
54
+ "normalized": false,
55
+ "rstrip": false,
56
+ "single_word": false,
57
+ "special": true
58
+ },
59
+ "248047": {
60
+ "content": "<|object_ref_start|>",
61
+ "lstrip": false,
62
+ "normalized": false,
63
+ "rstrip": false,
64
+ "single_word": false,
65
+ "special": true
66
+ },
67
+ "248048": {
68
+ "content": "<|object_ref_end|>",
69
+ "lstrip": false,
70
+ "normalized": false,
71
+ "rstrip": false,
72
+ "single_word": false,
73
+ "special": true
74
+ },
75
+ "248049": {
76
+ "content": "<|box_start|>",
77
+ "lstrip": false,
78
+ "normalized": false,
79
+ "rstrip": false,
80
+ "single_word": false,
81
+ "special": true
82
+ },
83
+ "248050": {
84
+ "content": "<|box_end|>",
85
+ "lstrip": false,
86
+ "normalized": false,
87
+ "rstrip": false,
88
+ "single_word": false,
89
+ "special": true
90
+ },
91
+ "248051": {
92
+ "content": "<|quad_start|>",
93
+ "lstrip": false,
94
+ "normalized": false,
95
+ "rstrip": false,
96
+ "single_word": false,
97
+ "special": true
98
+ },
99
+ "248052": {
100
+ "content": "<|quad_end|>",
101
+ "lstrip": false,
102
+ "normalized": false,
103
+ "rstrip": false,
104
+ "single_word": false,
105
+ "special": true
106
+ },
107
+ "248053": {
108
+ "content": "<|vision_start|>",
109
+ "lstrip": false,
110
+ "normalized": false,
111
+ "rstrip": false,
112
+ "single_word": false,
113
+ "special": true
114
+ },
115
+ "248054": {
116
+ "content": "<|vision_end|>",
117
+ "lstrip": false,
118
+ "normalized": false,
119
+ "rstrip": false,
120
+ "single_word": false,
121
+ "special": true
122
+ },
123
+ "248055": {
124
+ "content": "<|vision_pad|>",
125
+ "lstrip": false,
126
+ "normalized": false,
127
+ "rstrip": false,
128
+ "single_word": false,
129
+ "special": true
130
+ },
131
+ "248056": {
132
+ "content": "<|image_pad|>",
133
+ "lstrip": false,
134
+ "normalized": false,
135
+ "rstrip": false,
136
+ "single_word": false,
137
+ "special": true
138
+ },
139
+ "248057": {
140
+ "content": "<|video_pad|>",
141
+ "lstrip": false,
142
+ "normalized": false,
143
+ "rstrip": false,
144
+ "single_word": false,
145
+ "special": true
146
+ },
147
+ "248058": {
148
+ "content": "<tool_call>",
149
+ "lstrip": false,
150
+ "normalized": false,
151
+ "rstrip": false,
152
+ "single_word": false,
153
+ "special": false
154
+ },
155
+ "248059": {
156
+ "content": "</tool_call>",
157
+ "lstrip": false,
158
+ "normalized": false,
159
+ "rstrip": false,
160
+ "single_word": false,
161
+ "special": false
162
+ },
163
+ "248060": {
164
+ "content": "<|fim_prefix|>",
165
+ "lstrip": false,
166
+ "normalized": false,
167
+ "rstrip": false,
168
+ "single_word": false,
169
+ "special": false
170
+ },
171
+ "248061": {
172
+ "content": "<|fim_middle|>",
173
+ "lstrip": false,
174
+ "normalized": false,
175
+ "rstrip": false,
176
+ "single_word": false,
177
+ "special": false
178
+ },
179
+ "248062": {
180
+ "content": "<|fim_suffix|>",
181
+ "lstrip": false,
182
+ "normalized": false,
183
+ "rstrip": false,
184
+ "single_word": false,
185
+ "special": false
186
+ },
187
+ "248063": {
188
+ "content": "<|fim_pad|>",
189
+ "lstrip": false,
190
+ "normalized": false,
191
+ "rstrip": false,
192
+ "single_word": false,
193
+ "special": false
194
+ },
195
+ "248064": {
196
+ "content": "<|repo_name|>",
197
+ "lstrip": false,
198
+ "normalized": false,
199
+ "rstrip": false,
200
+ "single_word": false,
201
+ "special": false
202
+ },
203
+ "248065": {
204
+ "content": "<|file_sep|>",
205
+ "lstrip": false,
206
+ "normalized": false,
207
+ "rstrip": false,
208
+ "single_word": false,
209
+ "special": false
210
+ },
211
+ "248066": {
212
+ "content": "<tool_response>",
213
+ "lstrip": false,
214
+ "normalized": false,
215
+ "rstrip": false,
216
+ "single_word": false,
217
+ "special": false
218
+ },
219
+ "248067": {
220
+ "content": "</tool_response>",
221
+ "lstrip": false,
222
+ "normalized": false,
223
+ "rstrip": false,
224
+ "single_word": false,
225
+ "special": false
226
+ },
227
+ "248068": {
228
+ "content": "<think>",
229
+ "lstrip": false,
230
+ "normalized": false,
231
+ "rstrip": false,
232
+ "single_word": false,
233
+ "special": false
234
+ },
235
+ "248069": {
236
+ "content": "</think>",
237
+ "lstrip": false,
238
+ "normalized": false,
239
+ "rstrip": false,
240
+ "single_word": false,
241
+ "special": false
242
+ },
243
+ "248070": {
244
+ "content": "<|audio_start|>",
245
+ "lstrip": false,
246
+ "normalized": false,
247
+ "rstrip": false,
248
+ "single_word": false,
249
+ "special": true
250
+ },
251
+ "248071": {
252
+ "content": "<|audio_end|>",
253
+ "lstrip": false,
254
+ "normalized": false,
255
+ "rstrip": false,
256
+ "single_word": false,
257
+ "special": true
258
+ },
259
+ "248072": {
260
+ "content": "<tts_pad>",
261
+ "lstrip": false,
262
+ "normalized": false,
263
+ "rstrip": false,
264
+ "single_word": false,
265
+ "special": true
266
+ },
267
+ "248073": {
268
+ "content": "<tts_text_bos>",
269
+ "lstrip": false,
270
+ "normalized": false,
271
+ "rstrip": false,
272
+ "single_word": false,
273
+ "special": true
274
+ },
275
+ "248074": {
276
+ "content": "<tts_text_eod>",
277
+ "lstrip": false,
278
+ "normalized": false,
279
+ "rstrip": false,
280
+ "single_word": false,
281
+ "special": true
282
+ },
283
+ "248075": {
284
+ "content": "<tts_text_bos_single>",
285
+ "lstrip": false,
286
+ "normalized": false,
287
+ "rstrip": false,
288
+ "single_word": false,
289
+ "special": true
290
+ },
291
+ "248076": {
292
+ "content": "<|audio_pad|>",
293
+ "lstrip": false,
294
+ "normalized": false,
295
+ "rstrip": false,
296
+ "single_word": false,
297
+ "special": true
298
+ }
299
+ }
300
+ }