[Model] Support Qwen3.8-Flash-Next (#53896)

Signed-off-by: huanghaoyan.hhy <huanghaoyan.hhy@alibaba-inc.com>
Signed-off-by: andyluo7 <andy.luo@amd.com>
Signed-off-by: lishunyang12 <lishunyang03@gmail.com>
Signed-off-by: Harry Huang <huanghaoyan.hhy@alibaba-inc.com>
Signed-off-by: Thien Tran <gau.nernst@yahoo.com.sg>
Signed-off-by: Isotr0py <Isotr0py@outlook.com>
Signed-off-by: Lucas Wilkinson <lwilkins@redhat.com>
Signed-off-by: Yifan Qiao <yifanqiao@inferact.ai>
Signed-off-by: zjy0516 <riverclouds.zhu@qq.com>
Co-authored-by: zjy0516 <riverclouds.zhu@qq.com>
Co-authored-by: KungYork <firelikest@163.com>
Co-authored-by: Thien Tran <gau.nernst@yahoo.com.sg>
Co-authored-by: Yifan Qiao <yifanqiao@inferact.ai>
Co-authored-by: andyluo7 <43718156+andyluo7@users.noreply.github.com>
Co-authored-by: Isotr0py <mozf@inferact.ai>
Co-authored-by: andyluo7 <andy.luo@amd.com>
Co-authored-by: OpenAI Codex <noreply@openai.com>
Co-authored-by: lishunyang12 <lishunyang03@gmail.com>
Co-authored-by: Vadim Gimpelson <156319763+vadiklyutiy@users.noreply.github.com>
Co-authored-by: Isotr0py <Isotr0py@outlook.com>
Co-authored-by: Lucas Wilkinson <lwilkins@redhat.com>
Co-authored-by: Claude Opus 5 <noreply@anthropic.com>
Co-authored-by: OpenAI Codex <codex@openai.com>
This commit is contained in:
Harry Huang
2026-08-31 05:57:56 +00:00
committed by GitHub
co-authored by zjy0516 KungYork Thien Tran Yifan Qiao andyluo7 Isotr0py andyluo7 OpenAI Codex lishunyang12 Vadim Gimpelson Isotr0py Lucas Wilkinson Claude Opus 5 OpenAI Codex
parent 687db59744
commit e126687a9a
124 changed files with 19424 additions and 450 deletions
+8 -3
View File
@@ -57,13 +57,13 @@ EPS = 1e-6
class _TestGatedNorm:
def __init__(self, weight: torch.Tensor) -> None:
def __init__(self, weight: torch.Tensor, activation: str) -> None:
self.weight = weight
self.bias = None
self.eps = EPS
self.group_size = None
self.norm_before_gate = True
self.activation = "silu"
self.activation = activation
def __call__(self, x: torch.Tensor, z: torch.Tensor) -> torch.Tensor:
return rmsnorm_fn(
@@ -99,6 +99,7 @@ def _build_layer(
dt_bias: torch.Tensor,
conv_weight: torch.Tensor,
norm_weight: torch.Tensor,
output_gate_activation: str,
):
layer = types.SimpleNamespace(
prefix=PREFIX,
@@ -116,7 +117,7 @@ def _build_layer(
dt_bias=dt_bias,
conv1d=types.SimpleNamespace(weight=conv_weight, bias=None),
kv_cache=(conv_state, ssm_state),
norm=_TestGatedNorm(norm_weight),
norm=_TestGatedNorm(norm_weight, output_gate_activation),
layer_norm_epsilon=EPS,
gdn_decode_kernel="cuda",
)
@@ -250,12 +251,14 @@ def test_fused_forward_uses_packed_entrypoint() -> None:
pytest.param([128], [1], [-1], 0, id="pure-decode"),
],
)
@pytest.mark.parametrize("output_gate_activation", ["silu", "sigmoid"])
@torch.inference_mode()
def test_fused_model_path_matches_reference(
seq_lens: list[int],
query_lens: list[int],
draft_tokens: list[int],
expected_fused_calls: int,
output_gate_activation: str,
) -> None:
"""Fused MTP and its mixed/prefill/decode fallbacks match the reference."""
torch.manual_seed(1)
@@ -334,6 +337,7 @@ def test_fused_model_path_matches_reference(
dt_bias,
conv_weight,
norm_weight,
output_gate_activation,
)
reference_out = torch.zeros_like(output_gate)
with patch.object(
@@ -355,6 +359,7 @@ def test_fused_model_path_matches_reference(
dt_bias,
conv_weight,
norm_weight,
output_gate_activation,
)
context.no_compile_layers = {PREFIX: fused_layer}
fused_out = torch.zeros_like(output_gate)
@@ -366,7 +366,7 @@ def test_preprocess_fused_align_matches_scalar_bookkeeping(monkeypatch, token_bi
input_batch=scalar_case[1],
requests=scalar_case[2],
forward_context={},
mamba_state_copy_funcs=(),
mamba_state_copy_funcs={},
copy_bufs=scalar_copy_bufs,
)
@@ -384,7 +384,7 @@ def test_preprocess_fused_align_matches_scalar_bookkeeping(monkeypatch, token_bi
input_batch=fused_case[1],
requests=fused_case[2],
forward_context={},
mamba_state_copy_funcs=(),
mamba_state_copy_funcs={},
copy_bufs=fused_copy_bufs,
align_ctx=ctx,
)