Merge branch 'main' into patches

Signed-off-by: MOVZX <movzx@yahoo.com>
This commit is contained in:
2026-10-16 17:35:42 +07:00
347 changed files with 29531 additions and 3472 deletions
+1 -10
View File
@@ -14,7 +14,6 @@ from tqdm import tqdm
import vllm.envs as envs
from vllm.distributed.parallel_state import (
get_dp_group,
get_ep_group,
is_global_first_rank,
)
from vllm.forward_context import ForwardContext, override_forward_context
@@ -356,7 +355,6 @@ def _get_fp8_fp4_mega_moe_m_values(
dg = _import_deep_gemm()
gate, experts = module.gate, module.experts
uses_mega_gate = _mega_moe_uses_mega_gate(module)
num_max_tokens_per_rank = experts.get_symm_buffer().num_max_tokens_per_rank
if module.use_sequence_parallel:
max_tokens = cdiv(max_tokens, module.tp_size)
@@ -369,14 +367,7 @@ def _get_fp8_fp4_mega_moe_m_values(
if uses_mega_gate and hasattr(dg, "get_bf16_mega_gate_config")
else {}
)
block_m = dg.get_block_m_for_mega_moe(
get_ep_group().world_size,
experts.num_experts,
num_max_tokens_per_rank,
num_tokens,
experts.top_k,
"fp8xfp4",
)
block_m = experts.get_block_m(num_tokens)
# Keep the largest M: DeepSeek-V4 only calls bf16_mega_gate above a
# small token threshold, and a config's largest M is always above it.
m_values[(tuple(gate_config.items()), block_m)] = num_tokens
@@ -30,6 +30,7 @@ _DEEPSEEK_V4_SPARSE_MLA_BACKENDS = frozenset(
"ROCM_FLASHMLA_SPARSE_DSV4",
"FLASHMLA_SPARSE_DSV41",
"FLASHINFER_MLA_SPARSE_DSV41",
"ROCM_FLASHMLA_SPARSE_DSV41",
"DEEPSEEK_SPARSE_SWA",
}
)