Merge branch 'main' into nightly
This commit is contained in:
Daniel Han
commit
11dcf38761
3 changed files with 22 additions and 11 deletions
|
|
@ -23,11 +23,11 @@ All notebooks are **beginner friendly**! Add your dataset, click "Run All", and
|
|||
| Unsloth supports | Free Notebooks | Performance | Memory use |
|
||||
|-----------|---------|--------|----------|
|
||||
| **Llama 3 (8B)** | [▶️ Start for free](https://colab.research.google.com/drive/135ced7oHytdxu3N2DNe1Z0kqjyYIkDXp?usp=sharing) | 2x faster | 60% less |
|
||||
| **Mistral v0.3 (7B)** | [▶️ Start for free](https://colab.research.google.com/drive/1_yNCks4BTD5zOnjozppphh5GzMFaMKq_?usp=sharing) | 2.2x faster | 73% less |
|
||||
| **Mistral Nemo (12B)** | [▶️ Start for free](https://colab.research.google.com/drive/17d3U-CAIwzmbDRqbZ9NnpHxCkmXB6LZ0?usp=sharing) | 2x faster | 60% less |
|
||||
| **Gemma 2 (9B)** | [▶️ Start for free](https://colab.research.google.com/drive/1vIrqH5uYDQwsJ4-OO3DErvuv4pBgVwk4?usp=sharing) | 2x faster | 63% less |
|
||||
| **Phi-3 (mini)** | [▶️ Start for free](https://colab.research.google.com/drive/1lN6hPQveB_mHSnTOYifygFcrO8C1bxq4?usp=sharing) | 2x faster | 50% less |
|
||||
| **Phi-3 (medium)** | [▶️ Start for free](https://colab.research.google.com/drive/1hhdhBa1j_hsymiW9m-WzxQtgqTH_NHqi?usp=sharing) | 2x faster | 50% less |
|
||||
| **Ollama** | [▶️ Start for free](https://colab.research.google.com/drive/1WZDi7APtQ9VsvOrQSSC5DDtxq159j8iZ?usp=sharing) | 1.9x faster | 43% less |
|
||||
| **Mistral v0.3 (7B)** | [▶️ Start for free](https://colab.research.google.com/drive/1_yNCks4BTD5zOnjozppphh5GzMFaMKq_?usp=sharing) | 2.2x faster | 73% less |
|
||||
| **ORPO** | [▶️ Start for free](https://colab.research.google.com/drive/11t4njE3c4Lxl-07OD8lJSMKkfyJml3Tn?usp=sharing) | 1.9x faster | 43% less |
|
||||
| **DPO Zephyr** | [▶️ Start for free](https://colab.research.google.com/drive/15vttTpzzVXv_tJwEk-hIcQ0S9FcEWvwP?usp=sharing) | 1.9x faster | 43% less |
|
||||
| **TinyLlama** | [▶️ Start for free](https://colab.research.google.com/drive/1AZghoNBQaMDgWJpi4RbffGM1h6raLUj9?usp=sharing) | 3.9x faster | 74% less |
|
||||
|
|
@ -39,8 +39,9 @@ All notebooks are **beginner friendly**! Add your dataset, click "Run All", and
|
|||
- Click [here](https://github.com/unslothai/unsloth/wiki) for detailed documentation for Unsloth.
|
||||
|
||||
## 🦥 Unsloth.ai News
|
||||
- 📣 NEW! [Mistral Nemo-12b](https://colab.research.google.com/drive/17d3U-CAIwzmbDRqbZ9NnpHxCkmXB6LZ0?usp=sharing) both Base and Instruct now supported
|
||||
- 📣 NEW! [Gemma-2-9b](https://colab.research.google.com/drive/1vIrqH5uYDQwsJ4-OO3DErvuv4pBgVwk4?usp=sharing) and Gemma-2-27b now supported
|
||||
- 📣 UPDATE! [Phi-3 mini](https://colab.research.google.com/drive/1hhdhBa1j_hsymiW9m-WzxQtgqTH_NHqi?usp=sharing) model updated
|
||||
- 📣 UPDATE! [Phi-3 mini](https://colab.research.google.com/drive/1hhdhBa1j_hsymiW9m-WzxQtgqTH_NHqi?usp=sharing) model updated. [Phi-3 Medium](https://colab.research.google.com/drive/1hhdhBa1j_hsymiW9m-WzxQtgqTH_NHqi?usp=sharing) 2x faster finetuning.
|
||||
- 📣 NEW! Continued Pretraining [notebook](https://colab.research.google.com/drive/1tEd1FrOXWMnCU9UIvdYhs61tkxdMuKZu?usp=sharing) for other languages like Korean!
|
||||
- 📣 NEW! Qwen2 now works
|
||||
- 📣 [Mistral v0.3 Base](https://colab.research.google.com/drive/1_yNCks4BTD5zOnjozppphh5GzMFaMKq_?usp=sharing) and [Mistral v0.3 Instruct]
|
||||
|
|
|
|||
|
|
@ -210,22 +210,24 @@ class GemmaFixedRotaryEmbedding(torch.nn.Module):
|
|||
self.dim = dim
|
||||
self.max_position_embeddings = max_position_embeddings
|
||||
self.base = base
|
||||
# Dynamic RoPE we first set it to a max of 4 * 8192 tokens then we iteratively grow this
|
||||
self.current_rope_size = min(4 * 8192, self.max_position_embeddings)
|
||||
|
||||
# Build here to make `torch.jit.trace` work.
|
||||
self._set_cos_sin_cache(seq_len=max_position_embeddings, device=device, dtype=torch.get_default_dtype())
|
||||
self._set_cos_sin_cache(seq_len=self.current_rope_size, device=device, dtype=torch.get_default_dtype())
|
||||
pass
|
||||
|
||||
def _set_cos_sin_cache(self, seq_len, device, dtype):
|
||||
# Note: on the original Llama codebase, these tensors are created on the target device (and not on CPU) and
|
||||
# in FP32. They are applied (multiplied) in FP32 as well.
|
||||
self.max_seq_len_cached = seq_len
|
||||
self.current_rope_size = seq_len
|
||||
|
||||
# The difference is we do division explicity instead of t * (1/x) ie we do t/x.
|
||||
freq_exponents = (2.0 / self.dim) * (
|
||||
torch.arange(self.dim // 2, dtype = torch.int64, device = "cpu").float()
|
||||
)
|
||||
timescale = self.base**freq_exponents
|
||||
positions = torch.arange(self.max_seq_len_cached, device = "cpu", dtype = torch.int64).float()
|
||||
positions = torch.arange(self.current_rope_size, device = "cpu", dtype = torch.int64).float()
|
||||
radians_new = positions[..., None] / timescale[None, None, :]
|
||||
radians_new = radians_new.squeeze(0)
|
||||
|
||||
|
|
@ -239,7 +241,7 @@ class GemmaFixedRotaryEmbedding(torch.nn.Module):
|
|||
|
||||
def forward(self, x, position_ids=None, seq_len=None):
|
||||
# x: [bs, num_attention_heads, seq_len, head_size]
|
||||
if seq_len > self.max_seq_len_cached:
|
||||
if seq_len > self.current_rope_size:
|
||||
self._set_cos_sin_cache(seq_len=seq_len, device=x.device, dtype=x.dtype)
|
||||
|
||||
return (
|
||||
|
|
@ -247,6 +249,13 @@ class GemmaFixedRotaryEmbedding(torch.nn.Module):
|
|||
self.sin_cached[:seq_len].to(dtype=x.dtype),
|
||||
)
|
||||
pass
|
||||
|
||||
def extend_rope_embedding(self, x, seq_len):
|
||||
if seq_len <= self.current_rope_size: return
|
||||
# Iteratively grow by increments of 8192
|
||||
self.current_rope_size = int(round(seq_len / 8192)) * 8192
|
||||
self._set_cos_sin_cache(self.current_rope_size, device = "cuda:0", dtype = x.dtype)
|
||||
pass
|
||||
pass
|
||||
|
||||
|
||||
|
|
@ -263,14 +272,14 @@ class GemmaFixedLinearScalingRotaryEmbedding(GemmaFixedRotaryEmbedding):
|
|||
def _set_cos_sin_cache(self, seq_len, device, dtype):
|
||||
# Note: on the original Llama codebase, these tensors are created on the target device (and not on CPU) and
|
||||
# in FP32. They are applied (multiplied) in FP32 as well.
|
||||
self.max_seq_len_cached = seq_len
|
||||
self.current_rope_size = seq_len
|
||||
|
||||
# The difference is we do division explicity instead of t * (1/x) ie we do t/x.
|
||||
freq_exponents = (2.0 / self.dim) * (
|
||||
torch.arange(self.dim // 2, dtype = torch.int64, device = "cpu").float()
|
||||
)
|
||||
timescale = self.base**freq_exponents
|
||||
positions = torch.arange(self.max_seq_len_cached, device = "cpu", dtype = torch.int64).float()
|
||||
positions = torch.arange(self.current_rope_size, device = "cpu", dtype = torch.int64).float()
|
||||
positions = positions / self.scaling_factor
|
||||
radians_new = positions[..., None] / timescale[None, None, :]
|
||||
radians_new = radians_new.squeeze(0)
|
||||
|
|
|
|||
|
|
@ -299,8 +299,9 @@ class FastMistralModel(FastLlamaModel):
|
|||
attention_module = MistralAttention,
|
||||
)
|
||||
# Just for Mistral Nemo models!
|
||||
function = patch_mistral_nemo_attention(function)
|
||||
if True:#init_name is not None:
|
||||
if function is not None:
|
||||
function = patch_mistral_nemo_attention(function)
|
||||
# if True:#init_name is not None:
|
||||
exec(function, globals())
|
||||
MistralAttention.__init__ = eval(init_name)
|
||||
pass
|
||||
|
|
|
|||
Loading…
Add table
Add a link
Reference in a new issue