[Feature][PCP] Add DeepSeek-V4.1-Flash support (#59857)
Signed-off-by: wangyicong <wangyicong@bytedance.com> Co-authored-by: OpenAI Codex <noreply@openai.com> Co-authored-by: Lucas Wilkinson <lwilkins@redhat.com> Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com> Co-authored-by: Kevin H. Luu <khluu000@gmail.com>
This commit is contained in:
co-authored by
OpenAI Codex
Lucas Wilkinson
Claude Opus 5.5
Kevin H. Luu
parent
074c83a09a
commit
fa75d3456d
@@ -0,0 +1,66 @@
|
||||
# SPDX-License-Identifier: Apache-2.0
|
||||
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
||||
from types import SimpleNamespace
|
||||
from unittest.mock import Mock
|
||||
|
||||
import pytest
|
||||
import torch
|
||||
|
||||
from vllm.models.deepseek_v41.nvidia import model as model_module
|
||||
|
||||
|
||||
@pytest.mark.parametrize(
|
||||
"pcp_size,use_ep", [(1, False), (2, True), (2, False), (4, False)]
|
||||
)
|
||||
@pytest.mark.parametrize("use_sp", [False, True])
|
||||
def test_pcp_gathers_routing_ids_once_after_local_input_preparation(
|
||||
pcp_size, use_ep, use_sp, monkeypatch
|
||||
):
|
||||
"""Embedding keeps local IDs; all decoder layers receive aligned routing IDs."""
|
||||
gather = Mock(side_effect=lambda ids, dim: torch.cat([ids] * pcp_size, dim=dim))
|
||||
monkeypatch.setattr(
|
||||
model_module, "get_pcp_group", lambda: SimpleNamespace(all_gather=gather)
|
||||
)
|
||||
monkeypatch.setattr(
|
||||
model_module,
|
||||
"get_pp_group",
|
||||
lambda: SimpleNamespace(is_first_rank=True, is_last_rank=True),
|
||||
)
|
||||
monkeypatch.setattr(model_module, "is_forward_context_available", lambda: False)
|
||||
monkeypatch.setattr(model_module, "sp_shard", lambda tensor: tensor[::2])
|
||||
monkeypatch.setattr(
|
||||
model_module, "sp_all_gather", lambda tensor: tensor.repeat_interleave(2, dim=0)
|
||||
)
|
||||
monkeypatch.setattr(model_module, "hc_collapse_triton", lambda hidden, mix: hidden)
|
||||
embed = Mock(side_effect=lambda ids: ids[:, None].float())
|
||||
run_layers = Mock(
|
||||
side_effect=lambda layer_ids, hidden, *args: (hidden, None, None, None, hidden)
|
||||
)
|
||||
model = SimpleNamespace(
|
||||
parallel_config=SimpleNamespace(
|
||||
prefill_context_parallel_size=pcp_size, enable_expert_parallel=use_ep
|
||||
),
|
||||
use_native_mega_moe=False,
|
||||
use_sequence_parallel=use_sp,
|
||||
engram_hash=None,
|
||||
embed_input_ids=embed,
|
||||
start_layer=0,
|
||||
decoder_replay_start=2,
|
||||
decoder_replay_layers=None,
|
||||
aux_hidden_state_layers=(),
|
||||
_mtp_hidden_buffer=None,
|
||||
_run_layers=run_layers,
|
||||
_collapse=lambda hidden, *args: hidden,
|
||||
norm=lambda hidden: hidden,
|
||||
)
|
||||
for offset in (0, 10):
|
||||
local_ids = torch.tensor([3, 7, 11, 15]) + offset
|
||||
model_module.DeepseekV4Model.forward(model, local_ids, torch.arange(4), None)
|
||||
assert embed.call_args.args[0] is local_ids
|
||||
routing_ids = local_ids[::2] if use_sp else local_ids
|
||||
needs_gather = pcp_size > 1 and not use_ep
|
||||
if needs_gather:
|
||||
torch.testing.assert_close(gather.call_args.args[0], routing_ids)
|
||||
routing_ids = routing_ids.repeat(pcp_size)
|
||||
torch.testing.assert_close(run_layers.call_args.args[3], routing_ids)
|
||||
assert gather.call_count == 2 * int(needs_gather)
|
||||
Reference in New Issue
Block a user