replace KEY_FULLATTN_BLK_IDX with KEY_WIN_ATTN_PATTERN

examples/llava/clip-impl.h

@@ -46,7 +46,7 @@
 #define KEY_MM_PATCH_MERGE_TYPE   "clip.vision.mm_patch_merge_type"
 #define KEY_IMAGE_GRID_PINPOINTS  "clip.vision.image_grid_pinpoints"
 #define KEY_IMAGE_CROP_RESOLUTION "clip.vision.image_crop_resolution"
-#define KEY_FULLATTN_BLK_IDX      "clip.vision.fullatt_block_indexes"
+#define KEY_WIN_ATTN_PATTERN      "clip.vision.n_wa_pattern"
 #define KEY_ATTN_WINDOW_SIZE      "clip.vision.window_size"
 
 

examples/llava/clip.cpp

@@ -171,7 +171,7 @@ struct clip_hparams {
     int32_t image_crop_resolution;
     std::unordered_set<int32_t> vision_feature_layer;
     int32_t attn_window_size;
-    std::vector<int32_t> full_attn_layers;
+    int32_t n_wa_pattern;
 };
 
 struct clip_layer {
@@ -799,7 +799,8 @@ static ggml_cgraph * clip_image_build_graph_qwen2_5_vl(clip_ctx * ctx, const cli
     const int n_head               = hparams.n_head;
     const int d_head               = hidden_size / n_head;
     const float eps                = hparams.eps;
-    const bool use_window_attn     = hparams.full_attn_layers.size() > 0;
+    const int n_wa_pattern         = hparams.n_wa_pattern;
+    const bool use_window_attn     = hparams.n_wa_pattern > 0;
     int mrope_sections[4] = {d_head/4, d_head/4, d_head/4, d_head/4};
 
     const int batch_size = imgs.entries.size();
@@ -926,8 +927,7 @@ static ggml_cgraph * clip_image_build_graph_qwen2_5_vl(clip_ctx * ctx, const cli
             V = ggml_reshape_3d(ctx0, V, num_positions, d_head, n_head * batch_size);
 
             struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q);
-            const bool inlist = std::find(hparams.full_attn_layers.begin(), hparams.full_attn_layers.end(), il) != hparams.full_attn_layers.end();
+            const bool full_attn = use_window_attn ? (il + 1) % n_wa_pattern == 0 : true;
-            const bool full_attn = use_window_attn ? inlist : true;
             if (full_attn) {
                 KQ = ggml_soft_max_ext(ctx0, KQ, nullptr, 1.0f / sqrtf((float)d_head), 0.0f);
             } else {
@@ -1721,8 +1721,8 @@ struct clip_model_loader {
             get_u32(KEY_PATCH_SIZE, hparams.patch_size);
             get_u32(KEY_IMAGE_CROP_RESOLUTION, hparams.image_crop_resolution, false);
             get_u32(KEY_ATTN_WINDOW_SIZE, hparams.attn_window_size, false);
+            get_u32(KEY_WIN_ATTN_PATTERN, hparams.n_wa_pattern, false);
             get_arr_int(KEY_IMAGE_GRID_PINPOINTS, hparams.image_grid_pinpoints, false);
-            get_arr_int(KEY_FULLATTN_BLK_IDX, hparams.full_attn_layers, false);
 
             {
                 std::string mm_patch_merge_type;
@@ -3074,6 +3074,7 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
     bool support_dynamic_size = ctx->has_minicpmv_projector
         || ctx->has_qwen2vl_merger
         || ctx->proj_type == PROJECTOR_TYPE_PIXTRAL;
+    const bool use_window_attn = hparams.n_wa_pattern > 0;
 
     const int image_size = hparams.image_size;
     int image_size_width  = image_size;
@@ -3335,7 +3336,7 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
         }
     }
 
-    if (hparams.attn_window_size > 0 && ctx->proj_type == PROJECTOR_TYPE_QWEN2_5_VL) {
+    if (use_window_attn && ctx->proj_type == PROJECTOR_TYPE_QWEN2_5_VL) {
         struct ggml_tensor * window_idx = ggml_graph_get_tensor(gf, "window_idx");
         struct ggml_tensor * inv_window_idx = ggml_graph_get_tensor(gf, "inv_window_idx");
         struct ggml_tensor * window_mask = ggml_graph_get_tensor(gf, "window_mask");
@@ -3388,9 +3389,9 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
             }
         }
 
-        if (window_idx) ggml_backend_tensor_set(window_idx, idx.data(), 0, ggml_nbytes(window_idx));
+        ggml_backend_tensor_set(window_idx, idx.data(), 0, ggml_nbytes(window_idx));
-        if (inv_window_idx) ggml_backend_tensor_set(inv_window_idx, inv_idx.data(), 0, ggml_nbytes(inv_window_idx));
+        ggml_backend_tensor_set(inv_window_idx, inv_idx.data(), 0, ggml_nbytes(inv_window_idx));
-        if (window_mask) ggml_backend_tensor_set(window_mask, mask.data(), 0, ggml_nbytes(window_mask));
+        ggml_backend_tensor_set(window_mask, mask.data(), 0, ggml_nbytes(window_mask));
     }
 
     ggml_backend_cpu_set_n_threads(ctx->backend_cpu, n_threads);

examples/llava/qwen2_vl_surgery.py

@@ -1,5 +1,5 @@
 import argparse
-from typing import Dict
+from typing import Dict, List, Optional
 
 import torch
 import numpy as np
@@ -20,6 +20,20 @@ VISION = "clip.vision"
 def k(raw_key: str, arch: str) -> str:
     return raw_key.format(arch=arch)
 
+
+def get_n_wa_pattern(fullatt_block_indexes: Optional[List[int]]):
+    if fullatt_block_indexes is None:
+        return 0
+    n_wa = fullatt_block_indexes[0]
+    for a, b in zip(fullatt_block_indexes, fullatt_block_indexes[1:]):
+        if b - a - 1 != n_wa:
+            raise ValueError(
+                f"window/full attention layer should have fix pattern of "
+                f"for each full-attention layer followed by {n_wa} window-attention layers"
+            )
+    return n_wa + 1
+
+
 class VL2:
 
     @staticmethod
@@ -152,7 +166,7 @@ def main(args):
         raise ValueError()
 
     if args.model_type == "qwen2.5vl":
-        fout.add_array("clip.vision.fullatt_block_indexes", vcfg.fullatt_block_indexes)
+        fout.add_uint32("clip.vision.n_wa_pattern", get_n_wa_pattern(vcfg.fullatt_block_indexes))
         fout.add_uint32("clip.vision.window_size", vcfg.window_size)
         fout.add_uint32(k(KEY_EMBEDDING_LENGTH, VISION), vcfg.hidden_size)
         fout.add_uint32("clip.vision.projection_dim", vcfg.out_hidden_size)