Comfy-Org/ComfyUI · error · ValueError

Unknown pos_emb_cls {self.pos_emb_cls}

Error message

Unknown pos_emb_cls {self.pos_emb_cls}

What it means

ComfyUI's Cosmos video DiT builds its positional-embedding module in build_pos_embed(), and the only supported class string is 'rope3d' (VideoRopePosition3DEmb). Any other value of the model config field pos_emb_cls raises ValueError at model construction. The field comes from the checkpoint's JSON config, so an unknown value almost always means an unsupported or mis-parsed Cosmos checkpoint.

Source

Thrown at comfy/ldm/cosmos/model.py:217

        else:
            self.affline_norm = nn.Identity()

        self.final_layer = FinalLayer(
            hidden_size=self.model_channels,
            spatial_patch_size=self.patch_spatial,
            temporal_patch_size=self.patch_temporal,
            out_channels=self.out_channels,
            use_adaln_lora=self.use_adaln_lora,
            adaln_lora_dim=self.adaln_lora_dim,
            weight_args=weight_args,
            operations=operations,
        )

    def build_pos_embed(self, device=None, dtype=None):
        if self.pos_emb_cls == "rope3d":
            cls_type = VideoRopePosition3DEmb
        else:
            raise ValueError(f"Unknown pos_emb_cls {self.pos_emb_cls}")

        logging.debug(f"Building positional embedding with {self.pos_emb_cls} class, impl {cls_type}")
        kwargs = dict(
            model_channels=self.model_channels,
            len_h=self.max_img_h // self.patch_spatial,
            len_w=self.max_img_w // self.patch_spatial,
            len_t=self.max_frames // self.patch_temporal,
            is_learnable=self.pos_emb_learnable,
            interpolation=self.pos_emb_interpolation,
            head_dim=self.model_channels // self.num_heads,
            h_extrapolation_ratio=self.rope_h_extrapolation_ratio,
            w_extrapolation_ratio=self.rope_w_extrapolation_ratio,
            t_extrapolation_ratio=self.rope_t_extrapolation_ratio,
            device=device,
        )
        self.pos_embedder = cls_type(
            **kwargs,
        )

View on GitHub (pinned to 1c6d8d45b3)

Solutions

1. Check the checkpoint's config JSON and set pos_emb_cls to "rope3d" (the only value this implementation supports).
2. If the checkpoint genuinely uses another embedding scheme, it is not supported by this port; use a Cosmos checkpoint that ships pos_emb_cls=rope3d.
3. Inspect how the config reaches the model (custom-node loader) and normalize/override pos_emb_cls before construction instead of editing the checkpoint.

Example fix

// before (config from checkpoint)
cfg = {"pos_emb_cls": "rope_3d", ...}
model = CosmosImageToVideoModel(cfg, ...)

// after
cfg = {**raw_cfg, "pos_emb_cls": "rope3d"}  # normalize to the only supported value
model = CosmosImageToVideoModel(cfg, ...)

SUPPORTED = {"rope3d"}
assert cfg.get("pos_emb_cls") in SUPPORTED, f"pos_emb_cls must be one of {SUPPORTED}, got {cfg.get('pos_emb_cls')!r}"

def is_valid_cosmos_pos_emb_cls(cfg: dict) -> bool:
    return cfg.get("pos_emb_cls") == "rope3d"

try:
    model.build_pos_embed(device, dtype)
except ValueError as e:
    raise RuntimeError(f"Unsupported Cosmos config: {e}; set pos_emb_cls='rope3d'") from e

When it happens

Trigger: Instantiating the Cosmos transformer with a config dict whose pos_emb_cls key is missing, misspelled (e.g. 'Rope3D', 'rope_3d'), or set to a variant (e.g. 'positional_embedding_3d') that this port does not implement. Happens when a custom node loads a Cosmos-family checkpoint with a config that was not remapped to ComfyUI's expected schema.

Common situations: Loading a new Cosmos/Predict2-variant checkpoint whose upstream config uses a different embedding name; a custom node passing raw model config through without sanitizing pos_emb_cls; typos in hand-written model config dicts.

Related errors

• Unknown interpolation method {self.interpolation}
• Unknown pos_emb_cls {self.pos_emb_cls}
• Normalization {name} not found
• Normalization mode {self.qkv_norm_mode} not found, only supp
• Unknown patch method: {self.patch_method}

AI-assisted analysis of Comfy-Org/ComfyUI@1c6d8d45b3 (2026-08-14). Data as JSON: /api/errors/6a921ea4722de14c. Report an issue: GitHub.