fix(mxfp4): Disable monolithic path for TRITON backend with EP (#34270)

Signed-off-by: Elizabeth Thomas <email2eliza@gmail.com>
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
Co-authored-by: Michael Goin <mgoin64@gmail.com>
This commit is contained in:
Elizabeth Thomas
2026-02-25 13:33:42 -08:00
committed by GitHub
co-authored by Claude Opus 4.6 Michael Goin
parent b188bab441
commit c97234c08b
2 changed files with 225 additions and 5 deletions
@@ -179,9 +179,35 @@ def triton_kernel_moe_forward(
global_num_experts: int = -1,
expert_map: torch.Tensor | None = None,
) -> torch.Tensor:
routing_data, gather_idx, scatter_idx = legacy_routing(
gating_output, topk, sm_first=not renormalize
)
if expert_map is not None:
# With expert parallelism, legacy_routing produces routing data
# using global expert IDs which don't correspond to local weight
# indices. Split the routing into topk selection + expert_map
# remapping + local routing data construction (matching the
# approach used by OAITritonExperts.apply).
from triton_kernels.topk import topk as topk_fn
sm_first = not renormalize
logits = gating_output
if sm_first:
logits = torch.softmax(logits, dim=-1)
sparse_logits = topk_fn(logits, topk, apply_softmax=not sm_first)
# sparse_logits.indx contains global expert IDs – remap to local.
topk_ids = expert_map[sparse_logits.indx.to(torch.long)]
topk_weights = sparse_logits.vals
local_num_experts = w1.size(0)
routing_data, gather_idx, scatter_idx = make_routing_data(
topk_ids, topk_weights, local_num_experts
)
# expert_map already applied; pass None downstream.
effective_expert_map = None
effective_global_num_experts = local_num_experts
else:
routing_data, gather_idx, scatter_idx = legacy_routing(
gating_output, topk, sm_first=not renormalize
)
effective_expert_map = expert_map
effective_global_num_experts = global_num_experts
output = torch.empty_like(hidden_states)
@@ -197,8 +223,8 @@ def triton_kernel_moe_forward(
activation=activation,
quant_config=quant_config,
apply_router_weight_on_input=apply_router_weight_on_input,
global_num_experts=global_num_experts,
expert_map=expert_map,
global_num_experts=effective_global_num_experts,
expert_map=effective_expert_map,
)