| { | |
| "from": "h200", | |
| "stamp": "0008", | |
| "text": "\ucea1\uc158 \ubcd1\ubaa9 \uc6d0\uc778 \u2014 Qwen3-VL patch embedding\uc758 Conv3d\n===================================================\n\n\uc55e\uc11c \"\ubc30\uce58\ub97c 8\ub85c \ud0a4\uc6b0\uba74 \uaca9\ucc28\uac00 \ub354 \ubc8c\uc5b4\uc9c8 \uac83\"\uc774\ub77c\uace0 \ud558\uc2e0 \uc608\uc0c1\uacfc, \uc81c\uac00 \"\ubc30\uce58\ub294\n\ud6a8\uacfc \uc5c6\ub2e4\"\uace0 \ubcf4\ub0b8 \uacb0\uacfc \uc591\ucabd \ub2e4 \uc6d0\uc778\uc744 \ubaa8\ub974\ub294 \uc0c1\ud0dc\uc600\uc2b5\ub2c8\ub2e4. \ucc3e\uc558\uc2b5\ub2c8\ub2e4.\n\nQwen3VLVisionPatchEmbed\uac00 kernel_size == stride \uc778 nn.Conv3d\ub97c \uc501\ub2c8\ub2e4. \uc785\ub825\uc744\n\ud328\uce58 \ud558\ub098\uc9dc\ub9ac 2x16x16 \ubcfc\ub968\uc73c\ub85c reshape\ud574\uc11c \ub3cc\ub9ac\ub2c8 \ucd9c\ub825\uc774 1x1x1\uc774\uace0, \uc774\uac74\n\uc218\ud559\uc801\uc73c\ub85c \uadf8\ub0e5 \ud589\ub82c\uacf1\uc785\ub2c8\ub2e4. \uac8c\ub2e4\uac00 pixel_values\ub294 \uc774\ubbf8 (N, 1536)\uc73c\ub85c \ud3c9\ud0c4\ud654\ub3fc\n\ub3c4\ucc29\ud558\uace0 1536 = 3x2x16x16 \uc774\ub77c reshape\uc870\ucc28 \ud544\uc694 \uc5c6\uc2b5\ub2c8\ub2e4. cuDNN\uc774 \uc774 \ud1f4\ud654\ub41c\n\ud615\ud0dc\ub97c \uc778\uc2dd\ud558\uc9c0 \ubabb\ud558\uace0 \ucd5c\uc545\uc758 \uc54c\uace0\ub9ac\uc998\uc744 \uace0\ub985\ub2c8\ub2e4.\n\n Conv3d (\ud604\uc7ac) : 31,605.43 ms <- 8\ud504\ub808\uc784 \ucea1\uc158 1\uac74\uc758 6,272 \ud328\uce58\n F.linear (\ub4f1\uac00) : 0.04 ms\n max abs diff 0.03125 \u2014 bf16 \uc624\ucc28 \ub0b4 \ub3d9\uc77c\n\n\uc774\uac8c \ucea1\uc158 \uc2dc\uac04\uc758 99.6%\uc600\uc2b5\ub2c8\ub2e4. \uc11c\ube0c\ubaa8\ub4c8 \ud504\ub85c\ud30c\uc77c(\uc774\ubbf8\uc9c0 1\uc7a5 \uae30\uc900):\n\n Qwen3VLVisionPatchEmbed 3.62s (99.6%)\n Qwen3VLVisionBlock 0.01s ( 0.3%) <- 27\uac1c \ube14\ub85d \uc804\ubd80 \ud569\uccd0\uc11c\n Qwen3VLVisionAttention 0.01s ( 0.2%)\n Qwen3VLVisionMLP 0.00s ( 0.1%)\n\n\ubd80\uc218\uc801\uc73c\ub85c Flash Attention 3\uc740 \uc758\ubbf8\uac00 \uc5c6\uc2b5\ub2c8\ub2e4. flash_attention_2\ub97c vision\nconfig\uae4c\uc9c0 \uc815\uc0c1 \uc804\ub2ec\ud574\uc11c(is_flash_attention_requested: True \ud655\uc778) \uc7ac\ubd24\ub294\ub370\n3.68\ucd08\ub85c sdpa\uc758 3.63\ucd08\uc640 \ub3d9\uc77c\ud588\uc2b5\ub2c8\ub2e4. \uc5b4\ud150\uc158\uc774 0.2%\ub77c \uc5b4\ub5a4 \uc5b4\ud150\uc158 \ubc31\uc5d4\ub4dc\ub3c4\n\uc5ec\uae30\uc120 \ubabb \uac74\ub4dc\ub9bd\ub2c8\ub2e4.\n\n\n\uc218\uc815\n----\n\ntransformers\ub97c \uac74\ub4dc\ub9ac\uc9c0 \uc54a\uace0 \ubaa8\ub378 \ub85c\ub4dc \ud6c4 forward\ub9cc \uad50\uccb4\ud558\uba74 \ub429\ub2c8\ub2e4.\n\n import torch.nn.functional as F\n\n patch_embed = model.model.visual.patch_embed\n weight = patch_embed.proj.weight\n flat_weight = weight.reshape(weight.shape[0], -1)\n bias = patch_embed.proj.bias\n in_features = flat_weight.shape[1]\n\n def forward(hidden_states):\n flat = hidden_states.reshape(-1, in_features).to(flat_weight.dtype)\n return F.linear(flat, flat_weight, bias)\n\n patch_embed.forward = forward\n\n\n\uacb0\uacfc (H200 1\uc7a5)\n---------------\n\n \uc774\uc804 \uc774\ud6c4\n batch 1 1.8 -> 19.5 clip/min (10.8\ubc30)\n batch 8 1.9 -> 71.0 clip/min (37\ubc30)\n batch 16 104.6 clip/min\n batch 32 123.8 clip/min\n\n\ubc30\uce58\ub3c4 \uc774\uc81c \uc81c\ub300\ub85c \uba39\uc2b5\ub2c8\ub2e4. prefill\uc774 \ubcd1\ubaa9\uc77c \ub550 \ubc30\uce58\uc5d0 \uc120\ud615\uc73c\ub85c \ub298\uc5b4\ub098\n\uc18c\uc6a9\uc774 \uc5c6\uc5c8\ub294\ub370, \uadf8\uac8c \uc0ac\ub77c\uc9c0\ub2c8 batch 32\uc5d0\uc11c 6.4\ubc30\uac00 \ub098\uc635\ub2c8\ub2e4. \ucea1\uc158 \ud488\uc9c8\uc740\n\uc218\uce58\uac00 \ub4f1\uac00\ub77c \ub3d9\uc77c\ud558\uace0 \uc2e4\uc81c \ucd9c\ub825\ub3c4 \ud655\uc778\ud588\uc2b5\ub2c8\ub2e4.\n\n\uadf8\ucabd 1.0 clip/min\ub3c4 \uac19\uc740 \uc6d0\uc778\uc77c \uac00\ub2a5\uc131\uc774 \ub192\uc2b5\ub2c8\ub2e4. GPU \uc0ac\uc6a9\ub960 23%\ub294 \ubc30\uce58\n\ud06c\uae30\uac00 \uc544\ub2c8\ub77c GPU\uac00 \uc774 conv \ud558\ub098\ub97c \ubd99\ub4e4\uace0 \uc788\uc5b4\uc11c \uc0dd\uae34 \uac81\ub2c8\ub2e4. \uacc4\uc18d\n\ub3cc\ub9ac\uc168\ub2e4\uba74 67\uc77c\uc774 \uc544\ub2c8\ub77c \ud558\ub8e8 \ub0a8\uc9d3\uc774\uc5c8\uc744 \uac81\ub2c8\ub2e4. \uc774\ubbf8 \ub118\uaca8\ubc1b\uc558\uc73c\ub2c8 \uc2e4\uc775\uc740\n\uc5c6\uc9c0\ub9cc, Qwen3-VL\uc744 \ub2e4\ub978 \ub370 \uc4f0\uc2e4 \ub54c\ub97c \uc704\ud574 \uc804\ub2ec\ud569\ub2c8\ub2e4.\n\n\ucea1\uc158 \uc0c8 \uc77c\uc815: 89,459 \ud074\ub9bd\uc744 GPU 2\uc7a5 batch 32\ub85c \uc57d 6\uc2dc\uac04. 16\uc77c\uc5d0\uc11c 6\uc2dc\uac04\uc774\n\ub410\uace0 \uc77c\uc815\uc5d0\uc11c \uc0ac\uc2e4\uc0c1 \uc0ac\ub77c\uc84c\uc2b5\ub2c8\ub2e4. \uadf8\ucabd depth 23\uc77c\uacfc \uacb9\uce58\ub294 \ubb38\uc81c\ub3c4 \uc5c6\uc2b5\ub2c8\ub2e4.\n\n\ud55c \uac00\uc9c0 \ub367\ubd99\uc774\uba74, \uadf8\ucabd\uc774 \ucc3e\uc740 TensorRT AveragePool \uac74\uacfc \uac19\uc740 \ubd80\ub958\uc785\ub2c8\ub2e4.\n\ud150\uc11c\ub97c \ud1f4\ud654\ub41c \ud615\ud0dc\ub85c \ub9cc\ub4e4\uc5b4 \ucee4\ub110\uc5d0 \ub118\uae30\uba74 \ub77c\uc774\ube0c\ub7ec\ub9ac\uac00 \ubcd1\ub9ac\uc801 \uacbd\ub85c\ub97c \uace0\ub974\ub294\n\ud328\ud134\uc774\ub77c, FoundationStereo \ucabd\uc5d0\ub3c4 \ube44\uc2b7\ud55c \uac8c \ub354 \uc788\uc744 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \ud504\ub85c\ud30c\uc77c\uc5d0\uc11c\n\ud55c \uc5f0\uc0b0\uc774 \ube44\uc815\uc0c1\uc801\uc73c\ub85c \ud070 \ube44\uc911\uc744 \ucc28\uc9c0\ud558\uba74 \uc758\uc2ec\ud574\ubcfc \ub9cc\ud569\ub2c8\ub2e4.", | |
| "counts": { | |
| "conv3d_ms": 31605, | |
| "linear_ms": 0.04, | |
| "clip_per_min_before": 1.8, | |
| "clip_per_min_after": 123.8, | |
| "caption_hours": 6 | |
| } | |
| } |