Llamas 3.1 405B fp4 changes upstreaming from 355_wip (#25135)

Signed-off-by: Aleksandr Malyshev <maleksan@amd.com>
Co-authored-by: Aleksandr Malyshev <maleksan@amd.com>
Co-authored-by: Doug Lehr <douglehr@amd.com>
This commit is contained in:
Aleksandr Malyshev
2025-09-25 19:16:53 -06:00
committed by GitHub
co-authored by Aleksandr Malyshev Doug Lehr
parent 8b77328ffe
commit 53a30845be
6 changed files with 301 additions and 38 deletions
@@ -8,6 +8,8 @@ import torch
from vllm.model_executor.custom_op import CustomOp
from .common import apply_rotary_emb_torch
from .rocm_aiter_rope_ops import (is_rocm_triton_rotary_embedding_enabled,
rocm_aiter_rotary_emb)
@CustomOp.register("rotary_embedding")
@@ -45,6 +47,8 @@ class RotaryEmbedding(CustomOp):
cache = cache.to(dtype)
self.cos_sin_cache: torch.Tensor
self.register_buffer("cos_sin_cache", cache, persistent=False)
self.is_rocm_triton_rotary_embedding_enabled = \
is_rocm_triton_rotary_embedding_enabled()
def _compute_inv_freq(self, base: float) -> torch.Tensor:
"""Compute the inverse frequency."""
@@ -120,14 +124,31 @@ class RotaryEmbedding(CustomOp):
return query, key
from vllm import _custom_ops as ops
self._match_cos_sin_cache_dtype(query)
# ops.rotary_embedding() is an in-place operation
# that updates the query and key tensors.
ops.rotary_embedding(positions, query, key, self.head_size,
self.cos_sin_cache, self.is_neox_style)
return query, key
def forward_hip(
self,
positions: torch.Tensor,
query: torch.Tensor,
key: Optional[torch.Tensor] = None,
) -> tuple[torch.Tensor, Optional[torch.Tensor]]:
if self.is_rocm_triton_rotary_embedding_enabled:
self._match_cos_sin_cache_dtype(query)
rocm_aiter_rotary_emb(positions, query, key, self.cos_sin_cache,
self.head_size, self.rotary_dim,
self.is_neox_style)
else:
# ops.rotary_embedding() is an in-place operation
# that updates the query and key tensors.
self.forward_cuda(positions, query, key)
return query, key
def forward_xpu(
self,
positions: torch.Tensor,
@@ -0,0 +1,86 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
import torch
import vllm.envs as envs
from vllm.platforms import current_platform
from vllm.utils import direct_register_custom_op
def is_rocm_triton_rotary_embedding_enabled() -> bool:
return (current_platform.is_rocm() and envs.VLLM_ROCM_USE_AITER
and envs.VLLM_ROCM_USE_TRITON_ROPE)
def rocm_aiter_rotary_emb_with_key_forward_triton_impl(
positions: torch.Tensor,
sin: torch.Tensor,
cos: torch.Tensor,
query: torch.Tensor,
key: torch.Tensor,
rotate_style: int = 0,
is_nope_first: bool = False,
) -> None:
import aiter.ops.triton.rope as ops
ops.rope_cached_thd_positions_2c_fwd_inplace(
query,
key,
cos,
sin,
positions,
rotate_style,
reuse_freqs_front_part=True,
nope_first=is_nope_first,
)
def rocm_aiter_rotary_emb_with_key_forward_triton_fake(
positions: torch.Tensor,
sin: torch.Tensor,
cos: torch.Tensor,
query: torch.Tensor,
key: torch.Tensor,
rotate_style: int = 0,
is_nope_first: bool = False,
) -> None:
pass
if is_rocm_triton_rotary_embedding_enabled():
direct_register_custom_op(
op_name="rocm_aiter_rotary_emb_with_key_forward_triton",
op_func=rocm_aiter_rotary_emb_with_key_forward_triton_impl,
mutates_args=["key", "query"],
fake_impl=rocm_aiter_rotary_emb_with_key_forward_triton_fake,
dispatch_key=current_platform.dispatch_key,
)
def rocm_aiter_rotary_emb(positions: torch.Tensor, query: torch.Tensor,
key: torch.Tensor, cos_sin_cache: torch.Tensor,
head_size: int, rotary_dim: int,
is_neox_style: bool):
num_tokens = positions.numel()
cos, sin = cos_sin_cache.chunk(2, dim=-1)
query_shape = query.shape
key_shape = key.shape
rotate_style = 0 if is_neox_style else 1
query = query.view(num_tokens, -1, head_size)
key = key.view(num_tokens, -1, head_size)
query_ = query[..., :rotary_dim]
key_ = key[..., :rotary_dim]
positions = positions.view(*query.shape[:1])
torch.ops.vllm.rocm_aiter_rotary_emb_with_key_forward_triton(
positions,
sin,
cos,
query_,
key_,
rotate_style,
False,
)
query = query.view(query_shape)
key = key.view(key_shape)