* Shark Studio SDXL support, HIP driver support, simpler device info, small fixes
* Fixups to llm API/UI and ignore user config files.
* Small fixes for unifying pipelines.
* Update requirements.txt for iree-turbine (#2130)
* Fix Llama2 on CPU (#2133)
* Filesystem cleanup and custom model fixes (#2127)
* Fix some formatting issues
* Remove IREE pin (fixes exe issue) (#2126)
* Update find links for IREE packages (#2136)
* Shark Studio SDXL support, HIP driver support, simpler device info, small fixes
* Abstract out SD pipelines from Studio Webui (WIP)
* Switch from pin to minimum torch version and fix index url
* Fix device parsing.
* Fix linux setup
* Fix custom weights.
---------
Co-authored-by: saienduri <77521230+saienduri@users.noreply.github.com>
Co-authored-by: gpetters-amd <159576198+gpetters-amd@users.noreply.github.com>
Co-authored-by: gpetters94 <gpetters@protonmail.com>
* (WIP): Studio2 app infra and SD API
UI/app structure and utility implementation.
- Initializers for webui/API launch
- Schedulers file for SD scheduling utilities
- Additions to API-level utilities
- Added embeddings module for LoRA, Lycoris, yada yada
- Added image_processing module for resamplers, resize tools,
transforms, and any image annotation (PNG metadata)
- shared_cmd_opts module -- sorry, this is stable_args.py. It lives on.
We still want to have some global control over the app exclusively
from the command-line. At least we will be free from shark_args.
- Moving around some utility pieces.
- Try to make api+webui concurrency possible in index.py
- SD UI -- this is just img2imgUI but hopefully a little better.
- UI utilities for your nod logos and your gradio temps.
Enable UI / bugfixes / tweaks
* Studio2/SD: Use more correct LoRA alpha calculation (#2034)
* Updates ProcessLoRA to use both embedded LoRA alpha, and lora_strength
optional parameter (default 1.0) when applying LoRA weights.
* Updates ProcessLoRA to cover more dim cases.
* This bring ProcessLoRA into line with PR #2015 against Studio1
* Studio2: Remove duplications from api/utils.py (#2035)
* Remove duplicate os import
* Remove duplicate parse_seed_input function
Migrating to JSON requests in SD UI
More UI and app flow improvements, logging, shared device cache
Model loading
Complete SD pipeline.
Tweaks to VAE, pipeline states
Pipeline tweaks, add cmd_opts parsing to sd api
* Add test for SD
* Small cleanup
* Shark2/SD/UI: Respect ckpt_dir, share and server_port args (#2070)
* Takes whether to generate a gradio live link from the existing --share command
line parameter, rather than hardcoding as True.
* Takes server port from existing --server_port command line parameter, rather than
hardcoding as 11911.
* Default --ckpt_dir parameter to '../models'
* Use --ckpt_dir rather than hardcoding ../models as the base directory for
checkpoints, vae, and lora, etc
* Add a 'checkpoints' directory below --ckpt_dir to match ComfyUI folder structure.
Read custom_weights choices from there, and/or subfolders below there matching
the selected base model.
* Fix --ckpt_dir possibly not working correctly when an absolute rather than relative path
is specified.
* Relabel "Custom Weights" to "Custom Weights Checkpoint" in the UI
* Add StreamingLLM support to studio2 chat (#2060)
* Streaming LLM
* Update precision and add gpu support
* (studio2) Separate weights generation for quantization support
* Adapt prompt changes to studio flow
* Remove outdated flag from llm compile flags.
* (studio2) use turbine vmfbRunner
* tweaks to prompts
* Update CPU path and llm api test.
* Change device in test to cpu.
* Fixes to runner, device names, vmfb mgmt
* Use small test without external weights.
* HF-Reference LLM mode + Update test result to match latest Turbine. (#2080)
* HF-Reference LLM mode.
* Fixup test to match current output from Turbine.
* lint
* Fix test error message + Only initialize HF torch model when used.
* Remove redundant format_out change.
* Add rest API endpoint from LanguageModel API
* Add StreamingLLM support to studio2 chat (#2060)
* Streaming LLM
* Update precision and add gpu support
* (studio2) Separate weights generation for quantization support
* Adapt prompt changes to studio flow
* Remove outdated flag from llm compile flags.
* (studio2) use turbine vmfbRunner
* tweaks to prompts
* Update CPU path and llm api test.
* Change device in test to cpu.
* Fixes to runner, device names, vmfb mgmt
* Use small test without external weights.
* Formatting and init files.
* Remove unused import.
* Small fixes
* Studio2/SD/UI: Improve various parts of the UI for Stable Diffusion (#2074)
* Studio2/SD/UI: Improve various parts of the UI of Shark 2
* Update Gradio pin to 4.15.0.
* Port workarounds for Gradio >4.8.0 main container sizing from Shark 1.0.
* Move nod Logo out of the SD tab and onto the top right of the main tab bar.
* Set nod logo icon as the favicon (as current Shark 1.0).
* Create a tabbed right hand panel within the SD UI sized to the viewport height.
* Make Input Image tab 1 in the right hand panel.
* Make output images, generation log, and generation buttons, tab 2 in the
right hand panel
* Make config JSON display, with config load, save and clear, tab 3 in the
right hand panel
* Make gallery area of the Output tab take up all vertical space the other controls
on the tab do not.
* Tidy up the controls on the Config tab somewhat.
* Studio2/SD/UI: Reorganise inputs on Left Panel of SD tab
* Rename previously added Right Panel Output tab to 'Generate'.
* Move Batch Count, Batch Size, and Repeatable Seeds, off of Left Panel and onto 'Generate' Tab.
* On 'Generate' tab, rename 'Generate Image(s)' button to 'Start', and 'Stop Batch' button to 'Stop'. They are now below the Batch inputs on a Generate tab so don't need the specificity.
* Move Device, Low VRAM, and Precision inputs into their own 'Device Settings' Accordion control. (starts closed)
* Rename 'Custom Weights Checkpoint' to 'Checkpoint Weights'
* Move Checkpoint Weights, VAE Model, Standalone Lora Weights, and Embeddings Options controls, into their own 'Model Weights' Accordion control. (starts closed)
* Move Denoising Strength, and Resample Type controls into their own 'Input Image Processing' Accordion. (starts closed)
* Move any remaining controls in the 'Advanced Options' Accorion directly onto the left panel, and remove then Accordion.
* Enable the copy button for all text boxes on the SD tab.
* Add emoji/unicode glphs to all top level controls and Accordions on the SD Left Panel.
* Start with the 'Generate' as the initially selected tab in the SD Right Panel, working around Gradio issue #7805
* Tweaks to SD Right Tab Panel vertical height.
* Studio2/SD/UI: Sizing tweaks for Right Panel, and >1920 width
* Set height of right panel using vmin rather than vh, with explicit affordances
for fixed areas above and below.
* Port >1920 width Gradio >4.8 CSS workaround from Shark 1.0.
* Studio2/SD: Fix sd pipeline up to "Windows not supported" (#2082)
* Studio2/SD: Fix sd pipeline up to "Windows not supported"
A number of fixes to the SD pipeline as run from the UI, up until the point that dynamo
complains "Windows not yet supported for torch.compile".
* Remove separate install of iree-runtime and iree-compile in setup_venv.ps1, and rely on the
versions installed via the Turbine requirements.txt. Fixes#2063 for me.
* Replace any "None" strings with python None when pulling the config in the UI.
* Add 'hf_auth_token' param to api StableDiffusion class, defaulting to None, and then pass
that in to the various Models where it is required and wasn't already being done before.
* Fix clip custom_weight_params being passed to export_clip_model as "external_weight_file"
rather than "external_weights"
* Don't pass non-existing "custom_vae" parameter to the Turbine Vae Model, instead
pass custom_vae as the "hf_model_id" if it is set. (this may be wrong in the custom vae
cast, but stops the code *always* breaking).
* Studio2/SD/UI: Improve UI config None handling
* When populating the UI from a JSON Config set controls to "None" for null/None
values.
* When generating a JSON Config from the UI set props to null/None for controls
set to "None".
* Use null rather string 'None' in the default config
---------
Co-authored-by: Ean Garvey <87458719+monorimet@users.noreply.github.com>
* Studio2/SD/UI: Further sd ui pipeline fixes (#2091)
On Windows, this gets us all the way failing in iree compile of the with SD 2.1 base.
- Fix merge errors with sd right pane config UI tab.
- Remove non-requirement.txt install/build of torch/mlir/iree/SRT in setup_venv.ps1, fixing "torch.compile not supported on Windows" error.
- Fix gradio deprecation warning for `root=` FileExplorer kwarg.
- Comment out `precision` and `max_length` kwargs being passed to unet, as not yet supported on main Turbine branch. Avoids keyword argument error.
* Tweak compile-time flags for SD submodels.
* Small fixes to sd, pin mpmath
* Add pyinstaller spec and imports script.
* Fix the .exe (#2101)
* Fix _IREE_TARGET_MAP (#2103) (#2108)
- Change target passed to iree for vulkan from 'vulkan'
to 'vulkan-spriv', as 'vulkan' is not a valid value for
--iree-hal-target-backends with the current iree compiler.
Co-authored-by: Stefan Kapusniak <121311569+one-lithe-rune@users.noreply.github.com>
* Cleanup sd model map.
* Update dependencies.
* Studio2/SD/UI: Update gradio to 4.19.2 (sd-studio2) (#2097)
- Move pin for gradio from 4.15 -> 4.19.2 on the sd-studio2 branch
* fix formatting and disable explicit vulkan env settings.
---------
Co-authored-by: Stefan Kapusniak <121311569+one-lithe-rune@users.noreply.github.com>
Co-authored-by: Stanley Winata <68087699+raikonenfnu@users.noreply.github.com>
Co-authored-by: gpetters-amd <159576198+gpetters-amd@users.noreply.github.com>
Co-authored-by: gpetters94 <gpetters@protonmail.com>
* improved sharded performance and fixed issue with lmhead on rocm
* mmap shards + disable sharing of device arrays across devices
* fix device_idx for non-layer vmfbs
* fix time calc for sharded
---------
Co-authored-by: Elias Joseph <elias@nod-labs.com>
Co-authored-by: PhaneeshB <b.phaneesh@gmail.com>
* Fix some issues with defaults
Fixes to llama2 cpu compilation (turns off data tiling for old argmax
mode)
---------
Co-authored-by: Max Dawkins <max.dawkins@gmail.com>
* Update default CPU compilation flags.
c5a6cdc8dd52eb7e9b82
tweak CPU iree-compile flags to match upstream changes.
* Add an option for data tiling on SD models.
* Move clean_device_info to compile_utils
* Update compile_utils.py
* Fix .mlir writes for some user-level permissions
* Fix cases where full URI is given
* Fix conditionals.
* Fix device path handling in vulkan utils.
compile_str is always False in compile_module_to_flatbuffer since there
is a parameter 'model_name' before 'debug'.
This issue is relative to https://github.com/nod-ai/SHARK/pull/1863.
Then we can use mlir model buffer in RAM to run inference.
* Switch most compile flows to use ireec.compile_file.
* re-add input type to compile_str path.
* Check if mlir_module exists before checking if it's a path or pyobject.
* Fix some save_dir cases
Print a note ahead of a potentially long inactivity to set the right expectations.
Separately, we should add progress to the UI and make this loading faster.
-- Currently SHARK suggests that vmfb has been saved, while
that is not the case and no vmfb is generated.
This creates a misdirection for IR/vmfbs which are of larger
size.
-- This commit therefore fixes that misdirection.
Signed-off-by: Abhishek Varma <abhishek@nod-labs.com>
This allows to pass more arguemnts to the IREE compiler
Example:
python my-app.py --additional_compile_args="--mlir-pretty-debuginfo --mlir-timing"
Co-authored-by: Boian Petkantchin <boian@nod-labs.com>
-- This commit adds Scaled Dot Product Flash Attention's decomposition
in shark_importer.
-- It also updates `iree-flow-enable-data-tiling` to `iree-opt-data-tiling`.
Signed-off-by: Abhishek Varma <abhishek@nod-labs.com>
* WIP: MSVC ROCM support for SHARK Studio
* Make get_iree_rocm_args platform-agnostic.
* Update stable_args.py
* Update rocm arg handling in SD utils
* Guard quantization imports.
Co-authored-by: jam https://github.com/jammm
* Optimize device enumeration overhead and log details on long operations.
* Various fixes to add `@functools.cache` to what should be one time, expensive, device enumeration and setup activities. Cuts several seconds off of initialization on my machine.
* Add detailed tracing to actual invocations if they exceed a certain timeout or have an exception.
* Add detailed tracing to loading status.
* By default detail logging is only printed if an operation takes an excessive amount of time. All logging/timing can be printed by setting the variable `$env:SHARK_DETAIL_TRACE = "1"`
* Remove cache from unhashable functions
* Adding metal_utils for iree_utils
* Add patch for making compile API work for both MEGABYTE and MiniGPT4 (#1559)
-- It also modifies the mega_test.py script
Signed-off-by: Abhishek Varma <abhishek@nod-labs.com>
* [SD] Update unet in_channels API and add PIL metadata to spec. (#1560)
* Fix deprecation warning for unet config.
* Include PIL metadata instead of hidden imports in SD spec.
* Fixing iree-metal-target-platform
* adding metal to txt2img pipeline
* Fixing Copyright date
* removing debug prints
* black lint formating
* fixing device dump
---------
Signed-off-by: Abhishek Varma <abhishek@nod-labs.com>
Co-authored-by: Abhishek Varma <avarma094@gmail.com>
Co-authored-by: Ean Garvey <87458719+monorimet@users.noreply.github.com>
Co-authored-by: powderluv <powderluv@users.noreply.github.com>
* Do not hardcode the name of the VM module in get_iree_module
* Add example JAX MiniLM inference
---------
Co-authored-by: Boian Petkantchin <boian@nod-labs.com>
Example:
$ python my_app.py --device_allocator caching debug
This will wrap the device allocator with first caching allocator then
debug allocator.
$ python my_app.py --device_allocator caching
Only wrap with caching allocator.
Co-authored-by: Boian Petkantchin <boian@nod-labs.com>
