Fix/llm typo - #258
Open
hebangwen wants to merge 970 commits into
Open
Conversation
-update Permute multicore backend -implemnet RequantIntAxis core parrallel Change-Id: Ie0975f344ed130be750ff5591902e433860d578b
Change-Id: Iba1972c5d95028fccc7ca74f24db93fe08904bf4
…ecker -cmp_res displayed within braces -using "?" instead of "x" for failed results in intermediate cores Change-Id: Ib4b1e9960adc4d58d11086957d0cbcec4fdbd24e
calibrate the softmax output max prob range, insert two mul ops after softmax and scores*v matmul to rescale the softmax output for better accuracy. Change-Id: I5e50db81405f49f1e1bc002cd7da5f81ce4c3d59
- middle shape should be changed for permute and TopK Change-Id: Idaf5808d57f74cac09812aa71adc80fcef453f44
- update backend for bm1690/sg2380 Change-Id: I37170d21f3e8cc8ee7ae3d55d55b081d4ad3a2bf
- fix llm_analyse - support glm4.1v Change-Id: I190cce3a49b178296772b3794b765dde919391c7
- Added StructOptimizePass in lib/Dialect/Top/Transforms/StructOptimize.cpp - Added Added specific structural diagram optimization operations before shape inference - Added the adaptation for the top dialect in gen_rewriter_config.py - Add rules in gen_rewriter_config.py to enable the use of struct_optimize [number] in the model_transform stage, thereby enabling the activation of specific patterns Change-Id: I100b060078e9652affb3624147b27440f76e5ee6
- bug: cannot return from OpenMP region, two return in OpenMP region Change-Id: I2fb156e4193f40c6c668bc5285cf0ed64a9b0bd6
- status_str var may not used Change-Id: Ic5b88abc340748c0fd336088cc53040a525e7ae2
- PermuteMulconstSwap use nextOp->getResult(0).getUsers().begin(), error while nextOp is a two-result op(like topk) Change-Id: I96f762a3b28b50946750bd305a9ba5cdebf414cb
- MatMulWithSlice pattern split matmul support weight dtype other than f32 Change-Id: Ieb15e91c96d56145e149e170bf21064322a9c81f
…8443/#/c/149323/ - in gerrit 149323 weight_shape as llvm::ArrayRef<int64_t> shape changed, so use std::vector<T> instead Change-Id: Ia190ae3dcd070173beea1c6dc830517d1504d2d9
- hik still use deformable_attention and roiExtractor need to fix interface - roiExtractor add max_roi_num param for hik Change-Id: I2d206b7245276942e984935572a4b6db362bf39f
- using Conv & Reorder Operations to implement Merger + MatMul Change-Id: Id3636ab3c81c6d4264b4238c8183cfefa4934da4
- add identity and query_pos as input - support reference_points.w == 4 Change-Id: Ife4ea53da19cfc5c4a63dfd9c09c4ab014df90e2
there is no softmax in yolov5s Change-Id: Ia695c151f1b9a11647aa2da3d1656a17b38a577a
-GatherElements Change-Id: Ida9180f431537a89d973b4a84ffcddd7bdfd88e2
- use module::getElementType instead Change-Id: I22faec4a11060958e944ffacd1562f380a76ca6c
- only transpose op can be fused into matmul Change-Id: I6aada4b15e17023ec239521a516b6911b3a3336d
- LayerGroup add more search strategy to find better shape_secs - Allow user change configs of LayerGroup via LgConfig - Optimize Rope broadcast reload case in LayerGroup - Enable to use LgCostCache when getting final group and in LocalMemoryAllocationPass Change-Id: I498ed08bbe59dc9ab108d9ff58fbb17c118e9ca1
- set is_permute_optimize=true if pattern match success Change-Id: I5ce6df99dc6282f3b9cf2c79b8c6066aa74b78ff
- jira MLIR-632 - case: device tensor + host scalar, need to add Host2Device after ShapeProducer Change-Id: I9bca2f13bb8ea8049d6ba5e2899acb4cd5dbecbd
- moveUnaryPermute has bug when BinaryShiftOp's weight is 1x1 Change-Id: Ic81a41654a636107c66814a3cd92c4bb0bd3d9ba
- fix SliceOp BackwardH & BackwardW - fix can_be_group_small_c judgement for ReshapeOp Change-Id: I6ccd05fe331250085f5a6bfead2fb31c71bb767b
- new llm converter for llama3.2 Change-Id: I62add163b9e0cd8f4cc203074af71e457aafed7e
- test by: llm_convert.py -m /workspace/Qwen3-4B-AWQ -s 512 --quantize w4f16 -c bm1690 --out_dir qwen3_4b Change-Id: I85b9f58445b970261ade08060f80bb8bec5ff486
- it's unnecessary to duplicate onnx into workspace Change-Id: Idd91087e791fe8f490fe0771fbe703138ce6c1b6
… hook - Replace all Chinese comments with English equivalents - Improve code readability for international contributors - Translate by qwen3-8b using tools uploaded on https://github.com/Artlesbol/llm-comments-translator# - Add language check in pre-commit hook --------- Change-Id: If96045ef7a8f008412bbf57abfe11ad846c8fcdb Co-authored-by: CharlesXin <110083740+charlesxzb@users.noreply.github.com>
- print more info of llm analysis Change-Id: I5e7c2f6d3555df250dd4a02b6604b6ae12b883b3
- support int8, int4, f8e4m3, f4e2m1 in w4a16/w8a16 quant mode - support int8, int4, f8e4m3, f4e2m1 in qtable Change-Id: I0f60c5e4a6d35b258ecdcfc26184f38b1cd2085e
- if the op after input/before output is float, the input/output will keep float even quant_input/quant_output enabled - together with quant_input_int8/quant_output_int8, the input/output will be set int8/uint8 Change-Id: I4cb4a303e70d09f8564cf3fa95acab87b75f3fa3
- simplify custom_operator input when custom_mode in part_quantize Change-Id: I8987cbcee1167b7b7361291f76ff08d176197966
- remove sophon, sophgo from release pack Change-Id: I68898fa2fa290049fc1762da5c750e05376ab38f
- update lstm backend Change-Id: I0362187dcc2bd5f09d294be281fd04b12a290cdd
- fix is_scalar of add, sub, mul, div in shape_inference Change-Id: Id6a017fd04468f1a33466f4bb5847312dfe31d72
- Add test case for PixelNorm3 with three different input shapes - Implement numerical stability improvements - Add combined test case for multiple inputs Change-Id: I28014fe8c1ea3fca023b4964aa48c240259adc2c
- change id of runtime: 155471 - JIRA: https://jira.sophgo.com/browse/MLIR-665 - reuse runtime memory for different models Change-Id: Ifa5b2042f9286e593b9ce0a72ea58d1492e34144
- mobilenet and yolo Change-Id: If3ff4d2126af4812502a3db73ec0ee7a13373f28
- update model_zoo and libraries Change-Id: I9a8b7c6d5e4f3a2b1c0d9e8f7a6b5c4d3e2f1a0b
- change id of runtime: 155038 Change-Id: I35044aee36c8bb9ce62ca2393cc61baa410d7f4b
- and the quant of sub/subconst is not well designed and hard to handle, use float Change-Id: Ic08e3c06e788ff144cca476c0c087264477c3d62
- 32768 is too large for mars3, use LOCAL_BANK_SIZE Change-Id: If9eb61ae7ddbf61f6eaca66062d803e16735f111
- set group size in qtable by "op_name quant_mode group size" - only support bm1684x dynamic quantization Change-Id: I3dd93b7b0c2b884bcb27f41e19dc0ea065195964
- add topo-sort pass to adjust IRs' sequence Change-Id: I5c959ec48451eff5873ac39e65cfbfc89915745b
- cuda inference for vit Change-Id: I4fecf9b7fb9a40548fab823ee4fbd55f87b90f1b
- remain CastOp in onnx and add Int32 Lowering Change-Id: Ic0e23da85854ddfbdcf1d567e5718ecb91b76b27
- user judged should be in the group Change-Id: I0c31175e0f7473a0bf783b822d11a5e81e282784
- for swint Change-Id: I71b42d01318c21248bded339c939e709921f8636
- Interp miss quantize info if scale equals to 1 Change-Id: I3d44cc4c656c4ccc540a75d4fb03a723751d1362
- TpuLang deformable_attention and multi_scale_deformable_attention add grid_sampler_mode for GridSamplerOp Change-Id: I7f2926178551395e289e2dc82f48c46f483529d5
- mix per token/channel by setting group size = -1 in qtable - only support bm1684x tpu.mlir Change-Id: I27b0b8305d86b588ffdce33296b810d369ce256e
- add dynamic multi core process - update 1688/1684x, other backend will update by new commit Change-Id: I3991bfeefb7931f8a32f0e41a8b90ff63895ca5b
- update backend for 2260/2260e .. - refine multicore layergroup dynamic Change-Id: Ic994f18145da5ce250cbb4f64a8851f17c1e84ca
-cv184x miss init data when create subnet Change-Id: I7e0875ac62d01f21194e03555d89f371912ab849
- adjust multicore code Change-Id: I9b194f8d97fe918196a4ed46bad7e8a6ab945aa1
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
16 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
fix typo in llm converters.