Anurag1734/cuda-error-resolution-analysis
07
1[2 {3 "post_stream": {4 "posts": [5 {6 "id": 472337,7 "name": "偉杰 陳",8 "username": "xblackdavid",9 "avatar_template": "/user_avatar/discuss.pytorch.org/xblackdavid/{size}/77535_2.png",10 "created_at": "2025-06-30T13:42:59.276Z",11 "cooked": "<p>Hello everyone,</p>\n<p>I’m an ADAS engineer and not an AI major, nor did I graduate with an AI-related thesis, but my current work requires me to start utilizing AI technologies.</p>\n<p>My tasks currently involve Behavioral Cloning, Contrastive Learning, and Data Visualization Analysis. For model validation, I use metrics such as loss curve, Accuracy, Recall, and F1 Score to evaluate performance on the training, validation, and test sets. So far, I’ve managed to achieve results that align with some theoretical expectations.</p>\n<p>My current model architecture is relatively simple: it consists of an Encoder for static feature extraction (implemented with an MLP - Multi-Layer Perceptron), coupled with a Policy Head for dynamic feature capturing (GRU - Gated Recurrent Unit combined with a Linear layer and Softmax activation).</p>\n<p>Question on Transfer Learning and End-to-End Training Strategies<br>\nI have some questions regarding the application strategies for Transfer Learning and End-to-End Learning. My main concern isn’t about specific training issues, but rather, I’d like to ask for your insights on the best practices when training neural networks:</p>\n<p>Direct End-to-End Training: Would you recommend training end-to-end directly, either when starting with a completely new network or when the model hits a training bottleneck?</p>\n<p>Staged Training Strategy: Alternatively, would you suggest separating the Encoder and Policy Head? For instance, initially using Contrastive Learning to stabilize the Encoder, and then performing Transfer Learning to train the Policy Head?</p>\n<p>Flexible Adjustment Strategy: Or would you advise starting directly with end-to-end training, and if issues arise later, then disassembling the components to use Contrastive Learning or Data Visualization Analysis to adjust the Encoder, or to identify if the problem lies with the Dynamic Feature Capturing Policy Head?</p>\n<p>I’ve actually tried all these approaches myself and generally feel that it depends on the specific situation. However, since my internal colleagues and I have differing opinions, I’d appreciate hearing from all experienced professionals here.</p>\n<p>Thanks for your help!</p>",12 "post_number": 1,13 "post_type": 1,14 "posts_count": 1,15 "updated_at": "2025-06-30T13:42:59.276Z",16 "reply_count": 0,17 "reply_to_post_number": null,18 "quote_count": 0,19 "incoming_link_count": 5,20 "reads": 8,21 "readers_count": 7,22 "score": 26.6,23 "yours": false,24 "topic_id": 221166,25 "topic_slug": "transfer-learning-v-s-end-to-end-training",26 "display_username": "偉杰 陳",27 "primary_group_name": null,28 "flair_name": null,29 "flair_url": null,30 "flair_bg_color": null,31 "flair_color": null,32 "flair_group_id": null,33 "badges_granted": [],34 "version": 1,35 "can_edit": false,36 "can_delete": false,37 "can_recover": false,38 "can_see_hidden_post": false,39 "can_wiki": false,40 "read": true,41 "user_title": null,42 "bookmarked": false,43 "actions_summary": [],44 "moderator": false,45 "admin": false,46 "staff": false,47 "user_id": 84879,48 "hidden": false,49 "trust_level": 0,50 "deleted_at": null,51 "user_deleted": false,52 "edit_reason": null,53 "can_view_edit_history": true,54 "wiki": false,55 "post_url": "/t/transfer-learning-v-s-end-to-end-training/221166/1",56 "can_accept_answer": false,57 "can_unaccept_answer": false,58 "accepted_answer": false,59 "topic_accepted_answer": null,60 "can_vote": false61 }62 ],63 "stream": [64 47233765 ]66 },67 "timeline_lookup": [68 [69 1,70 11771 ]72 ],73 "suggested_topics": [74 {75 "fancy_title": "ImportError while installing mamba-ssm",76 "id": 215907,77 "title": "ImportError while installing mamba-ssm",78 "slug": "importerror-while-installing-mamba-ssm",79 "posts_count": 3,80 "reply_count": 1,81 "highest_post_number": 3,82 "image_url": null,83 "created_at": "2025-01-26T22:42:26.991Z",84 "last_posted_at": "2025-01-29T16:10:29.436Z",85 "bumped": true,86 "bumped_at": "2025-01-29T16:10:29.436Z",87 "archetype": "regular",88 "unseen": false,89 "pinned": false,90 "unpinned": null,91 "visible": true,92 "closed": false,93 "archived": false,94 "bookmarked": null,95 "liked": null,96 "tags_descriptions": {},97 "like_count": 1,98 "views": 1683,99 "category_id": 1,100 "featured_link": null,101 "has_accepted_answer": true,102 "posters": [103 {104 "extras": "latest",105 "description": "Original Poster, Most Recent Poster, Accepted Answer",106 "user": {107 "id": 78321,108 "username": "Omaralmaqtari",109 "name": "Omar al-maqtari",110 "avatar_template": "/user_avatar/discuss.pytorch.org/omaralmaqtari/{size}/72219_2.png",111 "trust_level": 1112 }113 },114 {115 "extras": null,116 "description": "Frequent Poster",117 "user": {118 "id": 3534,119 "username": "ptrblck",120 "name": "",121 "avatar_template": "/user_avatar/discuss.pytorch.org/ptrblck/{size}/1823_2.png",122 "admin": true,123 "moderator": true,124 "trust_level": 2125 }126 }127 ]128 },129 {130 "fancy_title": "How to Load Llama-3.3-70B-Instruct Model in Float8 Precision?",131 "id": 214102,132 "title": "How to Load Llama-3.3-70B-Instruct Model in Float8 Precision?",133 "slug": "how-to-load-llama-3-3-70b-instruct-model-in-float8-precision",134 "posts_count": 1,135 "reply_count": 0,136 "highest_post_number": 1,137 "image_url": null,138 "created_at": "2024-12-11T11:59:40.038Z",139 "last_posted_at": "2024-12-11T11:59:40.109Z",140 "bumped": true,141 "bumped_at": "2024-12-11T11:59:40.109Z",142 "archetype": "regular",143 "unseen": false,144 "pinned": false,145 "unpinned": null,146 "visible": true,147 "closed": false,148 "archived": false,149 "bookmarked": null,150 "liked": null,151 "tags_descriptions": {},152 "like_count": 0,153 "views": 160,154 "category_id": 1,155 "featured_link": null,156 "has_accepted_answer": false,157 "posters": [158 {159 "extras": "latest single",160 "description": "Original Poster, Most Recent Poster",161 "user": {162 "id": 78017,163 "username": "Arunima_Ghosh",164 "name": "Arunima Ghosh",165 "avatar_template": "/user_avatar/discuss.pytorch.org/arunima_ghosh/{size}/75750_2.png",166 "trust_level": 1167 }168 }169 ]170 },171 {172 "fancy_title": "How can I save the proper embeddings and weights after training?",173 "id": 212912,174 "title": "How can I save the proper embeddings and weights after training?",175 "slug": "how-can-i-save-the-proper-embeddings-and-weights-after-training",176 "posts_count": 2,177 "reply_count": 0,178 "highest_post_number": 2,179 "image_url": null,180 "created_at": "2024-11-13T08:45:07.754Z",181 "last_posted_at": "2024-11-13T12:37:00.417Z",182 "bumped": true,183 "bumped_at": "2024-11-13T15:11:25.864Z",184 "archetype": "regular",185 "unseen": false,186 "pinned": false,187 "unpinned": null,188 "visible": true,189 "closed": false,190 "archived": false,191 "bookmarked": null,192 "liked": null,193 "tags_descriptions": {},194 "like_count": 0,195 "views": 68,196 "category_id": 1,197 "featured_link": null,198 "has_accepted_answer": false,199 "posters": [200 {201 "extras": "latest single",202 "description": "Original Poster, Most Recent Poster",203 "user": {204 "id": 72736,205 "username": "songsong0425",206 "name": "Songyeon Lee",207 "avatar_template": "/user_avatar/discuss.pytorch.org/songsong0425/{size}/67200_2.png",208 "trust_level": 1209 }210 }211 ]212 },213 {214 "fancy_title": "Flex attention benchmarking",215 "id": 212440,216 "title": "Flex attention benchmarking",217 "slug": "flex-attention-benchmarking",218 "posts_count": 1,219 "reply_count": 0,220 "highest_post_number": 1,221 "image_url": null,222 "created_at": "2024-11-02T06:29:15.428Z",223 "last_posted_at": "2024-11-02T06:29:15.480Z",224 "bumped": true,225 "bumped_at": "2024-11-02T06:29:15.480Z",226 "archetype": "regular",227 "unseen": false,228 "pinned": false,229 "unpinned": null,230 "visible": true,231 "closed": false,232 "archived": false,233 "bookmarked": null,234 "liked": null,235 "tags_descriptions": {},236 "like_count": 0,237 "views": 193,238 "category_id": 1,239 "featured_link": null,240 "has_accepted_answer": false,241 "posters": [242 {243 "extras": "latest single",244 "description": "Original Poster, Most Recent Poster",245 "user": {246 "id": 76970,247 "username": "Lakshya_LNU",248 "name": "Lakshya LNU",249 "avatar_template": "/user_avatar/discuss.pytorch.org/lakshya_lnu/{size}/61585_2.png",250 "trust_level": 1251 }252 }253 ]254 },255 {256 "fancy_title": "Using autograd to calculate taylor expansions",257 "id": 216898,258 "title": "Using autograd to calculate taylor expansions",259 "slug": "using-autograd-to-calculate-taylor-expansions",260 "posts_count": 1,261 "reply_count": 0,262 "highest_post_number": 1,263 "image_url": null,264 "created_at": "2025-02-19T14:42:27.531Z",265 "last_posted_at": "2025-02-19T14:42:27.576Z",266 "bumped": true,267 "bumped_at": "2025-02-19T14:42:27.576Z",268 "archetype": "regular",269 "unseen": false,270 "pinned": false,271 "unpinned": null,272 "visible": true,273 "closed": false,274 "archived": false,275 "bookmarked": null,276 "liked": null,277 "tags_descriptions": {},278 "like_count": 0,279 "views": 67,280 "category_id": 1,281 "featured_link": null,282 "has_accepted_answer": false,283 "posters": [284 {285 "extras": "latest single",286 "description": "Original Poster, Most Recent Poster",287 "user": {288 "id": 11440,289 "username": "LeanderK",290 "name": "Leander Kurscheidt",291 "avatar_template": "/user_avatar/discuss.pytorch.org/leanderk/{size}/8099_2.png",292 "trust_level": 1293 }294 }295 ]296 }297 ],298 "tags_descriptions": {},299 "fancy_title": "Transfer learning v.s. end-to-end training",300 "id": 221166,301 "title": "Transfer learning v.s. end-to-end training",302 "posts_count": 1,303 "created_at": "2025-06-30T13:42:59.241Z",304 "views": 28,305 "reply_count": 0,306 "like_count": 0,307 "last_posted_at": "2025-06-30T13:42:59.276Z",308 "visible": true,309 "closed": false,310 "archived": false,311 "has_summary": false,312 "archetype": "regular",313 "slug": "transfer-learning-v-s-end-to-end-training",314 "category_id": 1,315 "word_count": 332,316 "deleted_at": null,317 "user_id": 84879,318 "featured_link": null,319 "pinned_globally": false,320 "pinned_at": null,321 "pinned_until": null,322 "image_url": null,323 "slow_mode_seconds": 0,324 "draft": null,325 "draft_key": "topic_221166",326 "draft_sequence": null,327 "unpinned": null,328 "pinned": false,329 "current_post_number": 1,330 "highest_post_number": 1,331 "deleted_by": null,332 "actions_summary": [333 {334 "id": 4,335 "count": 0,336 "hidden": false,337 "can_act": false338 },339 {340 "id": 8,341 "count": 0,342 "hidden": false,343 "can_act": false344 },345 {346 "id": 10,347 "count": 0,348 "hidden": false,349 "can_act": false350 },351 {352 "id": 7,353 "count": 0,354 "hidden": false,355 "can_act": false356 }357 ],358 "chunk_size": 20,359 "bookmarked": false,360 "topic_timer": null,361 "message_bus_last_id": 0,362 "participant_count": 1,363 "show_read_indicator": false,364 "thumbnails": null,365 "slow_mode_enabled_until": null,366 "can_vote": false,367 "vote_count": 0,368 "user_voted": false,369 "discourse_zendesk_plugin_zendesk_id": null,370 "discourse_zendesk_plugin_zendesk_url": "https://your-url.zendesk.com/agent/tickets/",371 "details": {372 "can_edit": false,373 "notification_level": 1,374 "participants": [375 {376 "id": 84879,377 "username": "xblackdavid",378 "name": "偉杰 陳",379 "avatar_template": "/user_avatar/discuss.pytorch.org/xblackdavid/{size}/77535_2.png",380 "post_count": 1,381 "primary_group_name": null,382 "flair_name": null,383 "flair_url": null,384 "flair_color": null,385 "flair_bg_color": null,386 "flair_group_id": null,387 "trust_level": 0388 }389 ],390 "created_by": {391 "id": 84879,392 "username": "xblackdavid",393 "name": "偉杰 陳",394 "avatar_template": "/user_avatar/discuss.pytorch.org/xblackdavid/{size}/77535_2.png"395 },396 "last_poster": {397 "id": 84879,398 "username": "xblackdavid",399 "name": "偉杰 陳",400 "avatar_template": "/user_avatar/discuss.pytorch.org/xblackdavid/{size}/77535_2.png"401 }402 },403 "bookmarks": []404 },405 {406 "post_stream": {407 "posts": [408 {409 "id": 471960,410 "name": "",411 "username": "bitzs",412 "avatar_template": "/letter_avatar_proxy/v4/letter/b/da6949/{size}.png",413 "created_at": "2025-06-19T13:35:50.533Z",414 "cooked": "<p>Hello everyone on the forum<br>\nIn my process of learning PyTorch quantization, I implemented convolution in two ways. Given quantized input and weight tensors, along with the scale and zero_point for the input, weight, and output tensors:</p>\n<p>The first approach involved converting the quantized input and weight tensors back to floating-point numbers, performing 2D convolution using the floating-point function <code>nn.Conv2d</code>, and then quantizing the result back to integers.</p>\n<p>The second approach directly used the <code>quantized.Conv2d</code> function to perform quantized convolution.</p>\n<p>After comparing the results from both methods, I noticed a considerable discrepancy. What could be the reason for this?</p>\n<pre><code class=\"lang-auto\">import torch.nn as nn\nimport torch.quantization\n\n\n# 量化参数设置 (示例值，根据实际需求修改)\ninput_scale = 0.012728\ninput_zero_point = 33\nweight_scale = 0.00292\nweight_zero_point = -11\nactivation_scale = 0.02091#0.02086#0.020389\nactivation_zero_point = 0\n\n\n# 读取32x32输入张量 (假设CSV是32行32列)#\ninput_tensor = torch.tensor([[[[200.,200.,200.,200.,200.],\n [200.,200.,200.,200.,200.],\n [200.,200.,200.,200.,200.],\n [200.,200.,200.,200.,200.],\n [200.,200.,200.,200.,200.]]]],dtype=torch.float32)\n# 读取5x5权重矩阵 (假设CSV是5行5列)\nweight_matrix =torch.tensor( [[[[ 9, -32, 17, 3, -67],\n [ -97, -41, 47, 84, -56],\n [-128, -125, 11, 63, 54],\n [ -15, -58, -31, 50, 48],\n [ 23, -3, 64, -87, -91]]]],dtype=torch.int8)\n\nbias = torch.tensor(0, dtype=torch.float32)\nbias = bias.reshape(-1)\n\nprint(input_tensor)\nprint(weight_matrix)\n\n#===========浮点模型计算=================\n# 创建浮点模型\nfloat_model = nn.Conv2d(\n in_channels=1,\n out_channels=1,\n kernel_size=5,\n stride=1,\n padding=0,\n bias=True\n )\n\nfloat_weight_matrix = weight_scale*(weight_matrix - weight_zero_point)\n\n# 手动设置权重\nwith torch.no_grad():\n float_model.weight = nn.Parameter(float_weight_matrix)\n float_model.bias = nn.Parameter(bias)\nfloat_input_tensor = input_scale*(input_tensor - input_zero_point)\n\nprint(float_model.weight)\nprint(float_input_tensor)\n\n# 执行量化卷积计算\nfloat_output = float_model(float_input_tensor)\n\nq_output = torch.quantize_per_tensor(\n float_output,\n scale= activation_scale,\n zero_point=activation_zero_point,\n dtype=torch.quint8\n)\n\nq_output_int = q_output.int_repr().squeeze().numpy()\n\n\n\n\n#=======================量化版本=========================\n\n#直接使用量化Conv2d\nquant_conv = torch.ao.nn.quantized.Conv2d(1, 1, 5, bias = False)\nquant_conv.scale = activation_scale\nquant_conv.zero_point = activation_zero_point\n\n\n# 量化权重张量\nquant_weight = torch.quantize_per_tensor(\n float_weight_matrix,\n scale=weight_scale,\n zero_point=weight_zero_point,\n dtype=torch.qint8\n)\n\n\n# 设置权重（使用浮点权重）\nquant_conv.set_weight_bias(quant_weight,None )\n\n\n# 量化输入张量\nquant_input = torch.quantize_per_tensor(\n float_input_tensor,\n scale=input_scale,\n zero_point=input_zero_point,\n dtype=torch.quint8\n)\n\n\nprint(\"Actual quantized weight:\", quant_conv.weight)\nprint(\"Integer representation:\", quant_conv.weight().int_repr())\n\n# 执行量化卷积\nquant_output = quant_conv(quant_input)\n\n\nprint(\"float_output float is:\", float_output)\nprint(\"float_output int is:\", q_output_int)\n\nprint(\"quant_output float is:\",quant_output)\nprint(\"quant_output int is:\",quant_output.int_repr())\nprint(\"compute done!\")\n</code></pre>\n<p>the result is<br>\n<div class=\"lightbox-wrapper\"><a class=\"lightbox\" href=\"https://discuss.pytorch.org/uploads/default/original/3X/f/0/f08b1bd84e46ac092e1a83b23aa03ee6a164134d.png\" data-download-href=\"https://discuss.pytorch.org/uploads/default/f08b1bd84e46ac092e1a83b23aa03ee6a164134d\" title=\"image\"><img src=\"https://discuss.pytorch.org/uploads/default/optimized/3X/f/0/f08b1bd84e46ac092e1a83b23aa03ee6a164134d_2_690x124.png\" alt=\"image\" data-base62-sha1=\"yjWDbGkEEJIlbSkkzeMC1XKnJr7\" width=\"690\" height=\"124\" srcset=\"https://discuss.pytorch.org/uploads/default/optimized/3X/f/0/f08b1bd84e46ac092e1a83b23aa03ee6a164134d_2_690x124.png, https://discuss.pytorch.org/uploads/default/original/3X/f/0/f08b1bd84e46ac092e1a83b23aa03ee6a164134d.png 1.5x, https://discuss.pytorch.org/uploads/default/original/3X/f/0/f08b1bd84e46ac092e1a83b23aa03ee6a164134d.png 2x\" data-dominant-color=\"292A2D\"><div class=\"meta\"><svg class=\"fa d-icon d-icon-far-image svg-icon\" aria-hidden=\"true\"><use href=\"#far-image\"></use></svg><span class=\"filename\">image</span><span class=\"informations\">865×156 79.3 KB</span><svg class=\"fa d-icon d-icon-discourse-expand svg-icon\" aria-hidden=\"true\"><use href=\"#discourse-expand\"></use></svg></div></a></div></p>",415 "post_number": 1,416 "post_type": 1,417 "posts_count": 3,418 "updated_at": "2025-06-19T13:35:50.533Z",419 "reply_count": 1,420 "reply_to_post_number": null,421 "quote_count": 0,422 "incoming_link_count": 4,423 "reads": 10,424 "readers_count": 9,425 "score": 17.0,426 "yours": false,427 "topic_id": 220920,428 "topic_slug": "why-is-there-such-a-significant-difference-between-floating-point-convolution-and-quantized-integer-convolution-results",429 "display_username": "",430 "primary_group_name": null,431 "flair_name": null,432 "flair_url": null,433 "flair_bg_color": null,434 "flair_color": null,435 "flair_group_id": null,436 "badges_granted": [],437 "version": 1,438 "can_edit": false,439 "can_delete": false,440 "can_recover": false,441 "can_see_hidden_post": false,442 "can_wiki": false,443 "read": true,444 "user_title": null,445 "bookmarked": false,446 "actions_summary": [],447 "moderator": false,448 "admin": false,449 "staff": false,450 "user_id": 84745,451 "hidden": false,452 "trust_level": 1,453 "deleted_at": null,454 "user_deleted": false,455 "edit_reason": null,456 "can_view_edit_history": true,457 "wiki": false,458 "post_url": "/t/why-is-there-such-a-significant-difference-between-floating-point-convolution-and-quantized-integer-convolution-results/220920/1",459 "can_accept_answer": false,460 "can_unaccept_answer": false,461 "accepted_answer": false,462 "topic_accepted_answer": null,463 "can_vote": false464 },465 {466 "id": 472323,467 "name": "Hamza Javaid",468 "username": "Hamza_Javaid",469 "avatar_template": "/user_avatar/discuss.pytorch.org/hamza_javaid/{size}/77517_2.png",470 "created_at": "2025-06-30T08:49:33.474Z",471 "cooked": "<aside class=\"quote no-group\" data-username=\"bitzs\" data-post=\"1\" data-topic=\"220920\">\n<div class=\"title\">\n<div class=\"quote-controls\"></div>\n<img alt=\"\" width=\"24\" height=\"24\" src=\"https://discuss.pytorch.org/letter_avatar_proxy/v4/letter/b/da6949/48.png\" class=\"avatar\"> bitzs:</div>\n<blockquote>\n<pre><code class=\"lang-auto\">import torch.nn as nn\nimport torch.quantization\n\n\n# 量化参数设置 (示例值，根据实际需求修改)\ninput_scale = 0.012728\ninput_zero_point = 33\nweight_scale = 0.00292\nweight_zero_point = -11\nactivation_scale = 0.02091#0.02086#0.020389\nactivation_zero_point = 0\n\n\n# 读取32x32输入张量 (假设CSV是32行32列)#\ninput_tensor = torch.tensor([[[[200.,200.,200.,200.,200.],\n [200.,200.,200.,200.,200.],\n [200.,200.,200.,200.,200.],\n [200.,200.,200.,200.,200.],\n [200.,200.,200.,200.,200.]]]],dtype=torch.float32)\n# 读取5x5权重矩阵 (假设CSV是5行5列)\nweight_matrix =torch.tensor( [[[[ 9, -32, 17, 3, -67],\n [ -97, -41, 47, 84, -56],\n [-128, -125, 11, 63, 54],\n [ -15, -58, -31, 50, 48],\n [ 23, -3, 64, -87, -91]]]],dtype=torch.int8)\n\nbias = torch.tensor(0, dtype=torch.float32)\nbias = bias.reshape(-1)\n</code></pre>\n</blockquote>\n</aside>\n<p>The crucial difference lies in how intermediate calculations are handled. When you perform convolution, you’re essentially doing many multiply-accumulate operations. In the quantized convolution implementation, these accumulations happen at higher precision (typically int32) before being rescaled to the output quantization parameters.</p>",472 "post_number": 2,473 "post_type": 1,474 "posts_count": 3,475 "updated_at": "2025-06-30T08:49:33.474Z",476 "reply_count": 0,477 "reply_to_post_number": null,478 "quote_count": 1,479 "incoming_link_count": 1,480 "reads": 9,481 "readers_count": 8,482 "score": 6.8,483 "yours": false,484 "topic_id": 220920,485 "topic_slug": "why-is-there-such-a-significant-difference-between-floating-point-convolution-and-quantized-integer-convolution-results",486 "display_username": "Hamza Javaid",487 "primary_group_name": null,488 "flair_name": null,489 "flair_url": null,490 "flair_bg_color": null,491 "flair_color": null,492 "flair_group_id": null,493 "badges_granted": [],494 "version": 1,495 "can_edit": false,496 "can_delete": false,497 "can_recover": false,498 "can_see_hidden_post": false,499 "can_wiki": false,500 "read": true,501 "user_title": null,502 "bookmarked": false,503 "actions_summary": [],504 "moderator": false,505 "admin": false,506 "staff": false,507 "user_id": 84865,508 "hidden": false,509 "trust_level": 2,510 "deleted_at": null,511 "user_deleted": false,512 "edit_reason": null,513 "can_view_edit_history": true,514 "wiki": false,515 "post_url": "/t/why-is-there-such-a-significant-difference-between-floating-point-convolution-and-quantized-integer-convolution-results/220920/2",516 "can_accept_answer": false,517 "can_unaccept_answer": false,518 "accepted_answer": false,519 "topic_accepted_answer": null520 },521 {522 "id": 472334,523 "name": "Jack Qasim",524 "username": "Jack_Qasim",525 "avatar_template": "/user_avatar/discuss.pytorch.org/jack_qasim/{size}/74623_2.png",526 "created_at": "2025-06-30T11:28:31.675Z",527 "cooked": "<p>Floating point convolution is more precise while quantized integer convolution reduces precision to save memory and speed up computation. The difference in precision can cause noticeable result differences especially in edge cases or small models.</p>",528 "post_number": 3,529 "post_type": 1,530 "posts_count": 3,531 "updated_at": "2025-06-30T11:28:31.675Z",532 "reply_count": 0,533 "reply_to_post_number": null,534 "quote_count": 0,535 "incoming_link_count": 0,536 "reads": 8,537 "readers_count": 7,538 "score": 1.6,539 "yours": false,540 "topic_id": 220920,541 "topic_slug": "why-is-there-such-a-significant-difference-between-floating-point-convolution-and-quantized-integer-convolution-results",542 "display_username": "Jack Qasim",543 "primary_group_name": null,544 "flair_name": null,545 "flair_url": null,546 "flair_bg_color": null,547 "flair_color": null,548 "flair_group_id": null,549 "badges_granted": [],550 "version": 1,551 "can_edit": false,552 "can_delete": false,553 "can_recover": false,554 "can_see_hidden_post": false,555 "can_wiki": false,556 "read": true,557 "user_title": null,558 "bookmarked": false,559 "actions_summary": [],560 "moderator": false,561 "admin": false,562 "staff": false,563 "user_id": 84884,564 "hidden": false,565 "trust_level": 0,566 "deleted_at": null,567 "user_deleted": false,568 "edit_reason": null,569 "can_view_edit_history": true,570 "wiki": false,571 "post_url": "/t/why-is-there-such-a-significant-difference-between-floating-point-convolution-and-quantized-integer-convolution-results/220920/3",572 "can_accept_answer": false,573 "can_unaccept_answer": false,574 "accepted_answer": false,575 "topic_accepted_answer": null576 }577 ],578 "stream": [579 471960,580 472323,581 472334582 ]583 },584 "timeline_lookup": [585 [586 1,587 128588 ],589 [590 2,591 117592 ]593 ],594 "suggested_topics": [595 {596 "fancy_title": "Right way to insert QuantStub and DeQuantStub in eager mode quantization",597 "id": 218666,598 "title": "Right way to insert QuantStub and DeQuantStub in eager mode quantization",599 "slug": "right-way-to-insert-quantstub-and-dequantstub-in-eager-mode-quantization",600 "posts_count": 7,601 "reply_count": 5,602 "highest_post_number": 7,603 "image_url": null,604 "created_at": "2025-04-06T15:53:04.268Z",605 "last_posted_at": "2025-04-12T02:36:54.395Z",606 "bumped": true,607 "bumped_at": "2025-04-12T02:36:54.395Z",608 "archetype": "regular",609 "unseen": false,610 "pinned": false,611 "unpinned": null,612 "visible": true,613 "closed": false,614 "archived": false,615 "bookmarked": null,616 "liked": null,617 "tags_descriptions": {},618 "like_count": 1,619 "views": 142,620 "category_id": 17,621 "featured_link": null,622 "has_accepted_answer": true,623 "posters": [624 {625 "extras": null,626 "description": "Original Poster",627 "user": {628 "id": 83665,629 "username": "Pepper9283",630 "name": null,631 "avatar_template": "/letter_avatar_proxy/v4/letter/p/848f3c/{size}.png",632 "trust_level": 1633 }634 },635 {636 "extras": "latest",637 "description": "Most Recent Poster, Accepted Answer",638 "user": {639 "id": 21770,640 "username": "jerryzh168",641 "name": "Jerry Zhang",642 "avatar_template": "/user_avatar/discuss.pytorch.org/jerryzh168/{size}/15217_2.png",643 "trust_level": 2644 }645 }646 ]647 },648 {649 "fancy_title": "Logits mismatch between PyTorch inference and manual implementation",650 "id": 219568,651 "title": "Logits mismatch between PyTorch inference and manual implementation",652 "slug": "logits-mismatch-between-pytorch-inference-and-manual-implementation",653 "posts_count": 2,654 "reply_count": 0,655 "highest_post_number": 2,656 "image_url": null,657 "created_at": "2025-04-28T18:57:49.995Z",658 "last_posted_at": "2025-04-29T00:42:05.572Z",659 "bumped": true,660 "bumped_at": "2025-04-29T00:42:05.572Z",661 "archetype": "regular",662 "unseen": false,663 "pinned": false,664 "unpinned": null,665 "visible": true,666 "closed": false,667 "archived": false,668 "bookmarked": null,669 "liked": null,670 "tags_descriptions": {},671 "like_count": 1,672 "views": 79,673 "category_id": 17,674 "featured_link": null,675 "has_accepted_answer": false,676 "posters": [677 {678 "extras": null,679 "description": "Original Poster",680 "user": {681 "id": 84057,682 "username": "greifswald",683 "name": "",684 "avatar_template": "/letter_avatar_proxy/v4/letter/g/90db22/{size}.png",685 "trust_level": 1686 }687 },688 {689 "extras": "latest",690 "description": "Most Recent Poster",691 "user": {692 "id": 21770,693 "username": "jerryzh168",694 "name": "Jerry Zhang",695 "avatar_template": "/user_avatar/discuss.pytorch.org/jerryzh168/{size}/15217_2.png",696 "trust_level": 2697 }698 }699 ]700 },701 {702 "fancy_title": "Torch.bfloat16 < how does it work? in bf 16 model",703 "id": 212450,704 "title": "Torch.bfloat16 < how does it work? in bf 16 model",705 "slug": "torch-bfloat16-how-does-it-work-in-bf-16-model",706 "posts_count": 2,707 "reply_count": 0,708 "highest_post_number": 2,709 "image_url": null,710 "created_at": "2024-11-02T14:14:26.984Z",711 "last_posted_at": "2024-11-04T17:14:31.044Z",712 "bumped": true,713 "bumped_at": "2024-11-04T17:14:31.044Z",714 "archetype": "regular",715 "unseen": false,716 "pinned": false,717 "unpinned": null,718 "visible": true,719 "closed": false,720 "archived": false,721 "bookmarked": null,722 "liked": null,723 "tags_descriptions": {},724 "like_count": 0,725 "views": 311,726 "category_id": 17,727 "featured_link": null,728 "has_accepted_answer": true,729 "posters": [730 {731 "extras": null,732 "description": "Original Poster",733 "user": {734 "id": 80013,735 "username": "minjun_jo",736 "name": "MJ_Jo",737 "avatar_template": "/user_avatar/discuss.pytorch.org/minjun_jo/{size}/63448_2.png",738 "trust_level": 1739 }740 },741 {742 "extras": "latest",743 "description": "Most Recent Poster, Accepted Answer",744 "user": {745 "id": 31938,746 "username": "Vasiliy_Kuznetsov",747 "name": "Vasiliy Kuznetsov",748 "avatar_template": "/user_avatar/discuss.pytorch.org/vasiliy_kuznetsov/{size}/24530_2.png",749 "trust_level": 2750 }751 }752 ]753 },754 {755 "fancy_title": "Is dynamic quantization in fact doing weight dequant instead of activation quant for `quantize_dynamic()`",756 "id": 217763,757 "title": "Is dynamic quantization in fact doing weight dequant instead of activation quant for `quantize_dynamic()`",758 "slug": "is-dynamic-quantization-in-fact-doing-weight-dequant-instead-of-activation-quant-for-quantize-dynamic",759 "posts_count": 2,760 "reply_count": 0,761 "highest_post_number": 2,762 "image_url": null,763 "created_at": "2025-03-12T21:55:04.470Z",764 "last_posted_at": "2025-03-13T12:56:28.178Z",765 "bumped": true,766 "bumped_at": "2025-03-13T12:56:28.178Z",767 "archetype": "regular",768 "unseen": false,769 "pinned": false,770 "unpinned": null,771 "visible": true,772 "closed": false,773 "archived": false,774 "bookmarked": null,775 "liked": null,776 "tags_descriptions": {},777 "like_count": 0,778 "views": 122,779 "category_id": 17,780 "featured_link": null,781 "has_accepted_answer": false,782 "posters": [783 {784 "extras": null,785 "description": "Original Poster",786 "user": {787 "id": 83226,788 "username": "Chiao-Wei_Hsu",789 "name": "Chiao-Wei Hsu",790 "avatar_template": "/user_avatar/discuss.pytorch.org/chiao-wei_hsu/{size}/76123_2.png",791 "trust_level": 0792 }793 },794 {795 "extras": "latest",796 "description": "Most Recent Poster",797 "user": {798 "id": 31938,799 "username": "Vasiliy_Kuznetsov",800 "name": "Vasiliy Kuznetsov",801 "avatar_template": "/user_avatar/discuss.pytorch.org/vasiliy_kuznetsov/{size}/24530_2.png",802 "trust_level": 2803 }804 }805 ]806 },807 {808 "fancy_title": "Quantization of Hybrid Pytorch Model",809 "id": 222931,810 "title": "Quantization of Hybrid Pytorch Model",811 "slug": "quantization-of-hybrid-pytorch-model",812 "posts_count": 1,813 "reply_count": 0,814 "highest_post_number": 1,815 "image_url": null,816 "created_at": "2025-09-08T04:47:22.142Z",817 "last_posted_at": "2025-09-08T04:47:22.197Z",818 "bumped": true,819 "bumped_at": "2025-09-08T04:47:22.197Z",820 "archetype": "regular",821 "unseen": false,822 "pinned": false,823 "unpinned": null,824 "visible": true,825 "closed": false,826 "archived": false,827 "bookmarked": null,828 "liked": null,829 "tags_descriptions": {},830 "like_count": 0,831 "views": 22,832 "category_id": 17,833 "featured_link": null,834 "has_accepted_answer": false,835 "posters": [836 {837 "extras": "latest single",838 "description": "Original Poster, Most Recent Poster",839 "user": {840 "id": 85546,841 "username": "Aman_230",842 "name": "",843 "avatar_template": "/user_avatar/discuss.pytorch.org/aman_230/{size}/74934_2.png",844 "trust_level": 1845 }846 }847 ]848 }849 ],850 "tags_descriptions": {},851 "fancy_title": "Why is there such a significant difference between floating-point convolution and quantized integer convolution results?",852 "id": 220920,853 "title": "Why is there such a significant difference between floating-point convolution and quantized integer convolution results?",854 "posts_count": 3,855 "created_at": "2025-06-19T13:35:50.498Z",856 "views": 47,857 "reply_count": 0,858 "like_count": 0,859 "last_posted_at": "2025-06-30T11:28:31.675Z",860 "visible": true,861 "closed": false,862 "archived": false,863 "has_summary": false,864 "archetype": "regular",865 "slug": "why-is-there-such-a-significant-difference-between-floating-point-convolution-and-quantized-integer-convolution-results",866 "category_id": 17,867 "word_count": 567,868 "deleted_at": null,869 "user_id": 84745,870 "featured_link": null,871 "pinned_globally": false,872 "pinned_at": null,873 "pinned_until": null,874 "image_url": "https://discuss.pytorch.org/uploads/default/original/3X/f/0/f08b1bd84e46ac092e1a83b23aa03ee6a164134d.png",875 "slow_mode_seconds": 0,876 "draft": null,877 "draft_key": "topic_220920",878 "draft_sequence": null,879 "unpinned": null,880 "pinned": false,881 "current_post_number": 1,882 "highest_post_number": 3,883 "deleted_by": null,884 "actions_summary": [885 {886 "id": 4,887 "count": 0,888 "hidden": false,889 "can_act": false890 },891 {892 "id": 8,893 "count": 0,894 "hidden": false,895 "can_act": false896 },897 {898 "id": 10,899 "count": 0,900 "hidden": false,901 "can_act": false902 },903 {904 "id": 7,905 "count": 0,906 "hidden": false,907 "can_act": false908 }909 ],910 "chunk_size": 20,911 "bookmarked": false,912 "topic_timer": null,913 "message_bus_last_id": 0,914 "participant_count": 3,915 "show_read_indicator": false,916 "thumbnails": [917 {918 "max_width": null,919 "max_height": null,920 "width": 865,921 "height": 156,922 "url": "https://discuss.pytorch.org/uploads/default/original/3X/f/0/f08b1bd84e46ac092e1a83b23aa03ee6a164134d.png"923 }924 ],925 "slow_mode_enabled_until": null,926 "can_vote": false,927 "vote_count": 0,928 "user_voted": false,929 "discourse_zendesk_plugin_zendesk_id": null,930 "discourse_zendesk_plugin_zendesk_url": "https://your-url.zendesk.com/agent/tickets/",931 "details": {932 "can_edit": false,933 "notification_level": 1,934 "participants": [935 {936 "id": 84745,937 "username": "bitzs",938 "name": "",939 "avatar_template": "/letter_avatar_proxy/v4/letter/b/da6949/{size}.png",940 "post_count": 1,941 "primary_group_name": null,942 "flair_name": null,943 "flair_url": null,944 "flair_color": null,945 "flair_bg_color": null,946 "flair_group_id": null,947 "trust_level": 1948 },949 {950 "id": 84865,951 "username": "Hamza_Javaid",952 "name": "Hamza Javaid",953 "avatar_template": "/user_avatar/discuss.pytorch.org/hamza_javaid/{size}/77517_2.png",954 "post_count": 1,955 "primary_group_name": null,956 "flair_name": null,957 "flair_url": null,958 "flair_color": null,959 "flair_bg_color": null,960 "flair_group_id": null,961 "trust_level": 2962 },963 {964 "id": 84884,965 "username": "Jack_Qasim",966 "name": "Jack Qasim",967 "avatar_template": "/user_avatar/discuss.pytorch.org/jack_qasim/{size}/74623_2.png",968 "post_count": 1,969 "primary_group_name": null,970 "flair_name": null,971 "flair_url": null,972 "flair_color": null,973 "flair_bg_color": null,974 "flair_group_id": null,975 "trust_level": 0976 }977 ],978 "created_by": {979 "id": 84745,980 "username": "bitzs",981 "name": "",982 "avatar_template": "/letter_avatar_proxy/v4/letter/b/da6949/{size}.png"983 },984 "last_poster": {985 "id": 84884,986 "username": "Jack_Qasim",987 "name": "Jack Qasim",988 "avatar_template": "/user_avatar/discuss.pytorch.org/jack_qasim/{size}/74623_2.png"989 }990 },991 "bookmarks": []992 },993 {994 "post_stream": {995 "posts": [996 {997 "id": 472238,998 "name": "Himalay Joriwal",999 "username": "Himalayjor",1000 "avatar_template": "/user_avatar/discuss.pytorch.org/himalayjor/{size}/49474_2.png",1001 "created_at": "2025-06-26T23:25:11.684Z",1002 "cooked": "<p>How can I disable guards generation when using torch.onnx.export(dynamo=True) with dynamic_shapes. Guards generated suggests fix which is wrong and doesn’t work.</p>\n<p>In my case, it suggests following fixes which is wrong. Model works with any number of frames in torchscript.</p>\n<pre><code class=\"lang-auto\">Suggested fixes:\n _frames = Dim('_frames', max=200)\n batch = 1\n frames = 16*_frames - 4\n</code></pre>",1003 "post_number": 1,1004 "post_type": 1,1005 "posts_count": 2,1006 "updated_at": "2025-06-26T23:25:11.684Z",1007 "reply_count": 1,1008 "reply_to_post_number": null,1009 "quote_count": 0,1010 "incoming_link_count": 26,1011 "reads": 10,1012 "readers_count": 9,1013 "score": 137.0,1014 "yours": false,1015 "topic_id": 221094,1016 "topic_slug": "disabling-guards-generation-using-dynamo-based-export",1017 "display_username": "Himalay Joriwal",1018 "primary_group_name": null,1019 "flair_name": null,1020 "flair_url": null,1021 "flair_bg_color": null,1022 "flair_color": null,1023 "flair_group_id": null,1024 "badges_granted": [],1025 "version": 1,1026 "can_edit": false,1027 "can_delete": false,1028 "can_recover": false,1029 "can_see_hidden_post": false,1030 "can_wiki": false,1031 "read": true,1032 "user_title": null,1033 "bookmarked": false,1034 "actions_summary": [],1035 "moderator": false,1036 "admin": false,1037 "staff": false,1038 "user_id": 55803,1039 "hidden": false,1040 "trust_level": 1,1041 "deleted_at": null,1042 "user_deleted": false,1043 "edit_reason": null,1044 "can_view_edit_history": true,1045 "wiki": false,1046 "post_url": "/t/disabling-guards-generation-using-dynamo-based-export/221094/1",1047 "can_accept_answer": false,1048 "can_unaccept_answer": false,1049 "accepted_answer": false,1050 "topic_accepted_answer": null,1051 "can_vote": false1052 },1053 {1054 "id": 472328,1055 "name": "Hamza Javaid",1056 "username": "Hamza_Javaid",1057 "avatar_template": "/user_avatar/discuss.pytorch.org/hamza_javaid/{size}/77517_2.png",1058 "created_at": "2025-06-30T09:05:39.315Z",1059 "cooked": "<aside class=\"quote no-group\" data-username=\"Himalayjor\" data-post=\"1\" data-topic=\"221094\">\n<div class=\"title\">\n<div class=\"quote-controls\"></div>\n<img alt=\"\" width=\"24\" height=\"24\" src=\"https://discuss.pytorch.org/user_avatar/discuss.pytorch.org/himalayjor/48/49474_2.png\" class=\"avatar\"> Himalayjor:</div>\n<blockquote>\n<pre><code class=\"lang-auto\">Suggested fixes:\n _frames = Dim('_frames', max=200)\n batch = 1\n frames = 16*_frames - 4\n</code></pre>\n</blockquote>\n</aside>\n<p>When you use <code>dynamo=True</code> in ONNX export, PyTorch tries to be helpful by suggesting constraints on your dynamic dimensions, but as you’ve discovered, these suggestions can be overly restrictive or simply incorrect.</p>\n<p>Try this first</p>\n<pre><code class=\"lang-auto\"># Reset dynamo state\ntorch._dynamo.reset()\n\n# Disable ALL shape specialization\ntorch._dynamo.config.automatic_dynamic_shapes = False\ntorch._dynamo.config.specialize_int = False\ntorch._dynamo.config.assume_static_by_default = False\ntorch._dynamo.config.guard_nn_modules = False\n\n# Now export your model\ntorch.onnx.export(\n model,\n example_input,\n \"model.onnx\",\n input_names=['input'],\n output_names=['output'],\n dynamic_axes={'input': {1: 'frames'}}, # Assuming frames is dimension 1\n dynamo=True,\n export_options=ExportOptions(dynamic_shapes=True)\n)\n</code></pre>\n<p>Also you can try a more legacy approach</p>\n<pre><code class=\"lang-auto\">torch.onnx.export(\n model,\n example_input,\n \"model.onnx\",\n input_names=['input'],\n output_names=['output'],\n dynamic_axes={'input': {1: 'frames'}},\n opset_version=17, # Use a recent opset\n # Note: no dynamo=True here\n)\n</code></pre>",1060 "post_number": 2,1061 "post_type": 1,1062 "posts_count": 2,1063 "updated_at": "2025-06-30T09:05:39.315Z",1064 "reply_count": 0,1065 "reply_to_post_number": null,1066 "quote_count": 1,1067 "incoming_link_count": 2,1068 "reads": 7,1069 "readers_count": 6,1070 "score": 11.4,1071 "yours": false,1072 "topic_id": 221094,1073 "topic_slug": "disabling-guards-generation-using-dynamo-based-export",1074 "display_username": "Hamza Javaid",1075 "primary_group_name": null,1076 "flair_name": null,1077 "flair_url": null,1078 "flair_bg_color": null,1079 "flair_color": null,1080 "flair_group_id": null,1081 "badges_granted": [],1082 "version": 1,1083 "can_edit": false,1084 "can_delete": false,1085 "can_recover": false,1086 "can_see_hidden_post": false,1087 "can_wiki": false,1088 "read": true,1089 "user_title": null,1090 "bookmarked": false,1091 "actions_summary": [],1092 "moderator": false,1093 "admin": false,1094 "staff": false,1095 "user_id": 84865,1096 "hidden": false,1097 "trust_level": 2,1098 "deleted_at": null,1099 "user_deleted": false,1100 "edit_reason": null,1101 "can_view_edit_history": true,1102 "wiki": false,1103 "post_url": "/t/disabling-guards-generation-using-dynamo-based-export/221094/2",1104 "can_accept_answer": false,1105 "can_unaccept_answer": false,1106 "accepted_answer": false,1107 "topic_accepted_answer": null1108 }1109 ],1110 "stream": [1111 472238,1112 4723281113 ]1114 },1115 "timeline_lookup": [1116 [1117 1,1118 1211119 ],1120 [1121 2,1122 1171123 ]1124 ],1125 "suggested_topics": [1126 {1127 "fancy_title": "How does torch dynamo set custom eval frame callback?",1128 "id": 219139,1129 "title": "How does torch dynamo set custom eval frame callback?",1130 "slug": "how-does-torch-dynamo-set-custom-eval-frame-callback",1131 "posts_count": 2,1132 "reply_count": 0,1133 "highest_post_number": 2,1134 "image_url": null,1135 "created_at": "2025-04-16T05:57:47.318Z",1136 "last_posted_at": "2025-04-16T12:09:20.702Z",1137 "bumped": true,1138 "bumped_at": "2025-04-16T12:09:20.702Z",1139 "archetype": "regular",1140 "unseen": false,1141 "pinned": false,1142 "unpinned": null,1143 "visible": true,1144 "closed": false,1145 "archived": false,1146 "bookmarked": null,1147 "liked": null,1148 "tags_descriptions": {},1149 "like_count": 0,1150 "views": 64,1151 "category_id": 39,1152 "featured_link": null,1153 "has_accepted_answer": false,1154 "posters": [1155 {1156 "extras": "latest single",1157 "description": "Original Poster, Most Recent Poster",1158 "user": {1159 "id": 68949,1160 "username": "huijjj",1161 "name": "HUIJONG JEONG",1162 "avatar_template": "/user_avatar/discuss.pytorch.org/huijjj/{size}/63435_2.png",1163 "trust_level": 11164 }1165 }1166 ]1167 },1168 {1169 "fancy_title": "Dynamic slicing torch.export",1170 "id": 212889,1171 "title": "Dynamic slicing torch.export",1172 "slug": "dynamic-slicing-torch-export",1173 "posts_count": 3,1174 "reply_count": 0,1175 "highest_post_number": 3,1176 "image_url": null,1177 "created_at": "2024-11-12T19:40:42.004Z",1178 "last_posted_at": "2024-11-16T17:48:29.488Z",1179 "bumped": true,1180 "bumped_at": "2024-11-16T17:48:29.488Z",1181 "archetype": "regular",1182 "unseen": false,1183 "pinned": false,1184 "unpinned": null,1185 "visible": true,1186 "closed": false,1187 "archived": false,1188 "bookmarked": null,1189 "liked": null,1190 "tags_descriptions": {},1191 "like_count": 0,1192 "views": 690,1193 "category_id": 39,1194 "featured_link": null,1195 "has_accepted_answer": false,1196 "posters": [1197 {1198 "extras": null,1199 "description": "Original Poster",1200 "user": {