xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2026-06-02 05:37:54 +08:00
Code Issues Packages Projects Releases Wiki Activity

cleanup deepep ll

Browse Source 
Signed-off-by: Sage Moore <sage@neuralmagic.com>
This commit is contained in:
Sage Moore 2025-07-03 13:35:21 +00:00
parent 6e2a3c0841
commit 17a7ceef27
1 changed files with 8 additions and 11 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

19
vllm/model_executor/layers/fused_moe/deepep_ll_prepare_finalize.py
View File

@ -131,8 +131,7 @@ class DeepEPLLPrepareAndFinalize(mk.FusedMoEPrepareAndFinalize):
hidden_size = a1.size(1) hidden_size = a1.size(1)
ubatch_ctx = get_current_ubatch_context() ubatch_ctx = get_current_ubatch_context()
ubatch_id = ubatch_ctx.id if ubatch_ctx is not None else -1 a2a_idx = ubatch_ctx.id if ubatch_ctx is not None else 0
a2a_idx = 0 if ubatch_id == -1 else ubatch_id
# assert hidden_size in self.SUPPORTED_HIDDEN_SIZES, \ # assert hidden_size in self.SUPPORTED_HIDDEN_SIZES, \
# (f"Hidden Size {hidden_size} not in supported list of hidden sizes" # (f"Hidden Size {hidden_size} not in supported list of hidden sizes"
# f"{self.SUPPORTED_HIDDEN_SIZES}") # f"{self.SUPPORTED_HIDDEN_SIZES}")
@ -155,8 +154,8 @@ class DeepEPLLPrepareAndFinalize(mk.FusedMoEPrepareAndFinalize):
a1 = a1 * rank_topk_weights.to(a1.dtype) a1 = a1 * rank_topk_weights.to(a1.dtype)
# Dispatch # Dispatch
# yield_and_switch_from_compute_to_comm_impl(schedule="default") yield_and_switch_from_compute_to_comm_impl(schedule="default")
expert_x, expert_num_tokens, handle, event, hook = \ expert_x, expert_num_tokens, handle, _, _= \
self.buffers[a2a_idx].low_latency_dispatch(a1, self.buffers[a2a_idx].low_latency_dispatch(a1,
rank_topk_ids, rank_topk_ids,
self.max_tokens_per_rank, self.max_tokens_per_rank,
@ -165,7 +164,7 @@ class DeepEPLLPrepareAndFinalize(mk.FusedMoEPrepareAndFinalize):
async_finish=False, async_finish=False,
return_recv_hook=False) return_recv_hook=False)
self.handles[a2a_idx] = handle self.handles[a2a_idx] = handle
# yield_and_switch_from_comm_to_compute_impl(schedule="default") yield_and_switch_from_comm_to_compute_impl(schedule="default")
expert_x, expert_x_scale = self._do_quant(expert_x, a1_scale, a2_scale, expert_x, expert_x_scale = self._do_quant(expert_x, a1_scale, a2_scale,
a1.dtype) a1.dtype)
@ -177,8 +176,7 @@ class DeepEPLLPrepareAndFinalize(mk.FusedMoEPrepareAndFinalize):
apply_router_weight_on_input: bool) -> None: apply_router_weight_on_input: bool) -> None:
ubatch_ctx = get_current_ubatch_context() ubatch_ctx = get_current_ubatch_context()
ubatch_id = ubatch_ctx.id if ubatch_ctx is not None else -1 a2a_idx = ubatch_ctx.id if ubatch_ctx is not None else 0
a2a_idx = 0 if ubatch_id == -1 else ubatch_id
handle = self.handles[a2a_idx] handle = self.handles[a2a_idx]
assert handle is not None assert handle is not None
@ -188,8 +186,8 @@ class DeepEPLLPrepareAndFinalize(mk.FusedMoEPrepareAndFinalize):
combine_topk_weights = torch.ones_like(topk_weights) combine_topk_weights = torch.ones_like(topk_weights)
# TODO (varun) : Enable zero copy mode # TODO (varun) : Enable zero copy mode
# yield_and_switch_from_compute_to_comm_impl(schedule="default") yield_and_switch_from_compute_to_comm_impl(schedule="default")
_, event, hook = self.buffers[a2a_idx].low_latency_combine( _ = self.buffers[a2a_idx].low_latency_combine(
fused_expert_output, fused_expert_output,
topk_ids, topk_ids,
combine_topk_weights, combine_topk_weights,
@ -198,6 +196,5 @@ class DeepEPLLPrepareAndFinalize(mk.FusedMoEPrepareAndFinalize):
zero_copy=False, zero_copy=False,
return_recv_hook=False, return_recv_hook=False,
out=output) out=output)
# event.current_stream_wait() yield_and_switch_from_comm_to_compute_impl(schedule="default")
# yield_and_switch_from_comm_to_compute_impl(schedule="default")