Commit fea46cb7 breaks Metal Stable Diffusion Generation

[super-fun-surf]

Commit fea46cb breaks the Image generation for the Metal pipeline.

still working commit 8696cf6

git checkout 8696cf64947a7f3b712297426078dcf6ab0d199e
Previous HEAD position was fea46cb7 Metal bgemm min changes (#2364)
HEAD is now at 8696cf64 Enable the affine kernel for u8/u32. (#2376)
➜  candle git:(8696cf64) cargo run --release --features metal  --example stable-diffusion -- --sd-version xl --n-steps 8 --width 1024 --height 1024 --use-f16 --prompt "crazy butterfly in the spring meadows, high quality, illustraion in the style of majoo mackinzee"
    Finished `release` profile [optimized] target(s) in 20.75s
     Running `target/release/examples/stable-diffusion --sd-version xl --n-steps 8 --width 1024 --height 1024 --use-f16 --prompt 'crazy butterfly in the spring meadows, high quality, illustraion in the style of majoo mackinzee'`
Running with prompt "crazy butterfly in the spring meadows, high quality, illustraion in the style of majoo mackinzee".
Building the Clip transformer.
Running with prompt "crazy butterfly in the spring meadows, high quality, illustraion in the style of majoo mackinzee".
Building the Clip transformer.

and it breaks on next commit fea46cb

➜  git checkout fea46cb719d5f59216f5b0a606400f1fd663190e
Previous HEAD position was bd80078a Fix log_sum_exp to handle large positive/negative inputs (#2367)
HEAD is now at fea46cb7 Metal bgemm min changes (#2364)
➜  candle git:(fea46cb7) cargo run --release --features metal  --example stable-diffusion -- --sd-version xl --n-steps 8 --width 1024 --height 1024 --use-f16 --prompt "crazy butterfly in the spring meadows, high quality, illustraion in the style of majoo mackinzee"
   Compiling candle-core v0.6.0 (/Users/pro/code/rust/candle/candle-core)
   Compiling candle-nn v0.6.0 (/Users/pro/code/rust/candle/candle-nn)
   Compiling candle-transformers v0.6.0 (/Users/pro/code/rust/candle/candle-transformers)
   Compiling candle-examples v0.6.0 (/Users/pro/code/rust/candle/candle-examples)
    Finished `release` profile [optimized] target(s) in 21.32s
     Running `target/release/examples/stable-diffusion --sd-version xl --n-steps 8 --width 1024 --height 1024 --use-f16 --prompt 'crazy butterfly in the spring meadows, high quality, illustraion in the style of majoo mackinzee'`
Running with prompt "crazy butterfly in the spring meadows, high quality, illustraion in the style of majoo mackinzee".
Building the Clip transformer.
Error: Metal error Error while loading function: "Compiler encountered an internal error"

[LaurentMazare]

Thanks for reporting this, tagging @ivarflakstad explicitly. If there is no easy fix I'll rollback the commit.

[ivarflakstad]

Interesting. What machine are you running this on?

[super-fun-surf]

Mac M1 16GB

[LaurentMazare]

We've reverted the problematic change in #2386 as @ivarflakstad has to conduct further investigations.