wonderful

关于测试浮点峰值的问题

2

 我现在跑的芯片型号是NVIDIA，ARMv8 Processor rev 0 (v8l)。 我看知乎文章里说测试浮点峰值时FMA指令的排布数量 = FMA的发射数 * FMA指令的延迟。我并没有查到上面这个芯片的手册。但是我看了A57的手册，里面是这样记录的：  FMA指令的延迟是10，吞吐量是2。我不太清楚这个吞吐是否代表着芯片可以同时发射两条FMA指令（是芯片发射吗），但是我分别放置了10条FMA指令（OP_FLOATS = 80）和20条FMA指令（OP_FLOATS = 160）都测试了，发现在10条的时候是16.095492 GFLOPS， 20条是 18.759214 GFLOPS。这是什么原因呢？ 我的猜测有两个： 1.10条FMA指令确实不是测试这款芯片的浮点峰值所需要的指令数。 2.可能编译器自动开启了多线程？这个比较有可能，因为从4条指令到10条指令性能差不多翻倍，但是10-20只增加了一点。

will runtime type checking go beyond function parameters and return type?

16

Great project, helps me understand DL code a lot. I used it like this: ``` python Patch_embed = Float32[torch.Tensor, f"B {PATCH_H} {PATCH_W} {PATCH_EMBED_DIM}"] Mlp_mid = Float32[torch.Tensor, f"B {PATCH_H} {PATCH_W} {MLP_HIDDEN}"]...

 It should be > (1 - ab_t[t, None, None, None]).sqrt() * noise according to the formula in the DDPM paper. 

Fix a 'bug' in prune_gaussians func

2

Change to prune Gaussians when there actually is any point need to remove. The old version will do remove_points even when to_remove is all false, which will discard all gradients...

Is habitat sim version v0.2.2 compitable?

5

Hi, I am encountering an issue while attempting to reproduce results reported in your paper by using the provided checkpoints. I cloned habitat-lab version 0.2.2 and habitat-sim version 0.2.2, and...

Bug: Why llamafile don't remove end token like <|eot_id|> or <end_of_turn>?

3

### Contact Details [email protected] ### What happened? When I use llamafile with python api. But for 2 models I use, they all retain the end token in response string, that...