diff --git a/vllm/envs.py b/vllm/envs.py index baa45b8fbd..4780fcc97f 100755 --- a/vllm/envs.py +++ b/vllm/envs.py @@ -395,7 +395,10 @@ if TYPE_CHECKING: # UVA backend is refused (vllm/v1/e2_guards.py). Off: the tree is upstream. # AOT: yes (different module class and forward). CAPTURE: yes (the captured # graph contains the wait/copy ops instead of a table lookup). Hashed. - VLLM_E2_PLE_PULL_TRANSPORT: bool = False + # UNSET = ON (since v2.5), the shipped value: upstream's UVA backend (the off + # path) compiles an fp8e4nv Triton kernel that sm_86 cannot build, so a launcher + # without this switch failed at profile_run. "0" still selects upstream's path. + VLLM_E2_PLE_PULL_TRANSPORT: bool = True # # RUNTIME-ONLY (registered for validate_environ + a single cached read): # Static K/V scale sidecar: fill_()ed into existing fp32 device buffers @@ -2518,7 +2521,7 @@ environment_variables: dict[str, Callable[[], Any]] = { "MERLIN_TOPK_EXACT_MEDIUM": lambda: os.getenv("MERLIN_TOPK_EXACT_MEDIUM", None), # PLE pull transport switch; strict "0"/"1" (see the TYPE_CHECKING block). "VLLM_E2_PLE_PULL_TRANSPORT": lambda: _e2_strict_switch( - "VLLM_E2_PLE_PULL_TRANSPORT" + "VLLM_E2_PLE_PULL_TRANSPORT", default=True ), # Runtime only; all of the following are in compile_factors.ignored_factors. "VLLM_E2_ALLOW_BREAKABLE": lambda: os.getenv("VLLM_E2_ALLOW_BREAKABLE", "") == "1", diff --git a/vllm/models/qwen4_exp/nvidia/ple_pull.py b/vllm/models/qwen4_exp/nvidia/ple_pull.py index f52000c8d9..25d9640b0c 100644 --- a/vllm/models/qwen4_exp/nvidia/ple_pull.py +++ b/vllm/models/qwen4_exp/nvidia/ple_pull.py @@ -15,8 +15,9 @@ class: ``Qwen4ExpNGramEmbeddingPull`` inherits it and ``PleOffloadLayer``. table is upstream's ``Qwen4ExpPLEDeviceEmbedding`` allocated on CPU; ``forward_impl`` computes n-gram IDs and gathers raw (FP8) rows. -Selected only when ``VLLM_E2_PLE_PULL_TRANSPORT=1``; with the switch off this -module is never imported by the model and the tree is upstream. +Selected when ``VLLM_E2_PLE_PULL_TRANSPORT`` is ``1`` or unset (the fork default +since v2.5); with ``=0`` this module is never imported by the model and the +tree is upstream. Supported table formats: FP8 (global scale) and unquantized -- the two ``Qwen4ExpPLEEmbeddingMethod`` families v0.30.0 itself supports. The E2 NVFP4