For tracing cond/while in eager, we trace the HOP with the eager backend with metadata torchfunction mode enabled. HOPs disallow the mutation that occurs in this torch function mode, so it is not able to be traced. As a result, we use a custom backend which enters this mode for tracing these HOPs. Thanks to @ydwu4 for the help with implementing this
Pull Request resolved: https://github.com/pytorch/pytorch/pull/134732
Approved by: https://github.com/ydwu4
For tracing cond/while in eager, we trace the HOP with the eager backend with metadata torchfunction mode enabled. HOPs disallow the mutation that occurs in this torch function mode, so it is not able to be traced. As a result, we use a custom backend which enters this mode for tracing these HOPs. Thanks to @ydwu4 for the help with implementing this
Pull Request resolved: https://github.com/pytorch/pytorch/pull/134732
Approved by: https://github.com/ydwu4
For tracing cond/while in eager, we trace the HOP with the eager backend with metadata torchfunction mode enabled. HOPs disallow the mutation that occurs in this torch function mode, so it is not able to be traced. As a result, we use a custom backend which enters this mode for tracing these HOPs. Thanks to @ydwu4 for the help with implementing this
Pull Request resolved: https://github.com/pytorch/pytorch/pull/134732
Approved by: https://github.com/ydwu4
This PR resolves#134408. Add an additional test and have passed the local test.
Do you think we should add a post-check to ensure `args` and `kwargs` are not both `None`? It seems to be possible to have modules without inputs.
This PR does not include any such post-check.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/134643
Approved by: https://github.com/zou3519
For tracing cond/while in eager, we trace the HOP with the eager backend with metadata torchfunction mode enabled. HOPs disallow the mutation that occurs in this torch function mode, so it is not able to be traced. As a result, we use a custom backend which enters this mode for tracing these HOPs. Thanks to @ydwu4 for the help with implementing this
Pull Request resolved: https://github.com/pytorch/pytorch/pull/134732
Approved by: https://github.com/ydwu4
…` and `attn_mask`, and correct device assignment for newly created variables in the method.
Fix example: Address broadcasting error in the addition of `attn_bias` and `attn_mask`, and correct device assignment for newly created variables in the method.
1. Adding `attn_bias += attn_mask` results in a broadcasting error. The expected shape of `attn_bias` is (L, S), so the output should also have the shape (L, S). However, when the input shape is (N, num_heads, L, S), broadcasting occurs, leading to an output shape of (N, num_heads, L, S), which is not desired.
2. `attn_bias` is a newly created variable within the method, but it is not assigned to the correct device.
**This is my retry of PR #130209 . The PR has been merged into commit `d4a79d4a7c746068d25fe5cf9333495561f4ce1f`, but the modifications were overwritten by subsequent commits.**
Co-authored-by: mikaylagawarecki <mikaylagawarecki@gmail.com>
@mikaylagawarecki provided a more elegant implementation.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/135427
Approved by: https://github.com/ezyang
The reraise is not supported and so this just gunks up our actual exception handling. You can trigger this by hitting an exception inside of an NN module that has hooks on it. You end up graph breaking on the reraise here, and losing the inner stack trace from the actual exception that was raised.
This might be kind of controversial. An alternate strategy is to support reraises in Dynamo or something but IDK this doesn't feel like the right place to apply force.
Signed-off-by: Edward Z. Yang <ezyang@meta.com>
Pull Request resolved: https://github.com/pytorch/pytorch/pull/133239
Approved by: https://github.com/anijain2305
As per title, this PR adds proper casting to fuse_linear_bn_weights in the same style as the conv case above. This previously caused numerical issues on my end, so that is why I am fixing it.
Also cleans up the docstring.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/134105
Approved by: https://github.com/mikaylagawarecki
### tl;dr
This PR adds GQA support to higher order op `flex_attention`.
## Details
When `enable_gqa` is set to True, HOP `flex_attention(score_mod, query, key, value, block_mask, enable_gqa)` runs Group Query Attention(GQA), where the number of query heads (Hq) is a multiple of number of key/value heads (Hkv). Each group of query heads (`Hq//Hkv` heads) attends to a shared kv head.
Otherwise, `flex_attention` assumes Multi Head Attention (MHA) where the number of query heads is equal the number of kv heads.
The `score_mod` and `mask_mod` API are adapted accordingly to take `q_head` as head index.
```
def score_mod(score: torch.Tensor, batch: torch.Tensor, q_head: torch.Tensor, token_q: torch.Tensor, token_kv: torch.Tensor) -> torch.Tensor
def mask_mod(batch: torch.Tensor, q_head: torch.Tensor, token_q: torch.Tensor, token_kv: torch.Tensor) -> torch.Tensor
```
## Example
```python
import torch
from torch.nn.attention.flex_attention import flex_attention
from torch.nn.attention.flex_attention import create_block_mask
torch.manual_seed(0)
def query_key_value_clones(
query: torch.Tensor,
key: torch.Tensor,
value: torch.Tensor,
dtype: torch.dtype = None,
):
"""Clones the query, key, and value tensors and moves them to the specified dtype."""
if dtype is None:
dtype = query.dtype
query_ref = query.clone().detach().to(dtype).requires_grad_(query.requires_grad)
key_ref = key.clone().detach().to(dtype).requires_grad_(key.requires_grad)
value_ref = value.clone().detach().to(dtype).requires_grad_(value.requires_grad)
return query_ref, key_ref, value_ref
# Lets create some input tensors
# The input tensor has shape (batch_size, num_heads, seq_len, head_dim).
# query and key/value can have different num_heads and seq_len
# Here 8 query heads share one KV head.
query = torch.randn(2, 8, 2048, 64, device="cuda", dtype=torch.float32, requires_grad=True)
key = torch.randn(2, 2, 2048, 64, device="cuda", dtype=torch.float32, requires_grad=True)
value = torch.randn(2, 2, 2048, 64, device="cuda", dtype=torch.float32, requires_grad=True)
query1, key1, value1 = query_key_value_clones(query, key, value)
# Lets create a score_modification. We take alibi_bias as an example.
# score_mod takes batch index, query head index, query index, and key/value index.
def _generate_alibi_bias(num_kv_heads: int, num_q_heads: int):
def _alibi_bias(
score: torch.Tensor,
b: torch.Tensor,
hq: torch.Tensor,
token_q: torch.Tensor,
token_kv: torch.Tensor,
) -> torch.Tensor:
# Let's calculate kv head from query head index
group = num_q_heads // num_kv_heads
hkv = hq // group
scale = torch.exp2(-((hkv + 1) * 8.0 / num_kv_heads))
return score + (token_kv - token_q) * scale
return _alibi_bias
# Let's apply a casual mask on top of it
def causal_mask(b, h, q, kv):
return q >= kv
# Generate a block mask for our new mask_mod function.
# The mask is broadcasted long head & batch dimensions.
block_mask = create_block_mask(causal_mask, B=1, H=1, Q_LEN=2048, KV_LEN=2048)
# Lets call flex_attention with our new score modification and block mask under eager mode.
output = flex_attention(query, key, value, score_mod=_generate_alibi_bias(2, 8), block_mask=block_mask, enable_gqa=True)
# Now lets compile flex_attention and run the flex_attention kernel.
compiled_flex_attention = torch.compile(flex_attention)
out_compiled = compiled_flex_attention(query1, key1, value1, score_mod=_generate_alibi_bias(2, 8), block_mask=block_mask, enable_gqa=True)
torch.testing.assert_close(output, out_compiled, atol=5e-2, rtol=2e-2)
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/131559
Approved by: https://github.com/drisspg
Overloads so that you can get more specific type info based on how you are indexing.
```python
from torch import nn
module_list = nn.ModuleList(32 * [nn.Linear(2, 2)])
# before:
reveal_type(module_list[0]) # Type of "module_list[0]" is "Module | ModuleList"
reveal_type(module_list[:1]) # Type of "module_list[: 1]" is "Module | ModuleList"
# now:
reveal_type(module_list[0]) # Type of "module_list[0]" is "Module"
reveal_type(module_list[:1]) # Type of "module_list[: 1]" is "ModuleList"
```
Co-authored-by: Skylion007 <Skylion007@users.noreply.github.com>
Pull Request resolved: https://github.com/pytorch/pytorch/pull/132834
Approved by: https://github.com/Skylion007, https://github.com/albanD
