fix: correct weight loading prefix mapping for Qwen3-VL

[Lollipop]

Summary

Fix Qwen3-VL-8B model producing garbage output due to incorrect weight loading.

Fixes #17887

Problem

The weight loading code unconditionally copies embed_tokens.weight to lm_head.weight:

if self.pp_group.is_last_rank and "model.embed_tokens.weight" in name:
    if "lm_head.weight" in params_dict:
        # copies embed_tokens to lm_head unconditionally

This is incorrect for models with tie_word_embeddings=False (like Qwen3-VL-8B), where lm_head has independent weights that should not be overwritten.

Model	tie_word_embeddings	lm_head weights
Qwen3-VL-2B	True	Shared with embed_tokens
Qwen3-VL-8B	False	Independent (should NOT be overwritten)

Fix

Add a check to only copy when tie_word_embeddings=True:

if (
    self.pp_group.is_last_rank
    and "model.embed_tokens.weight" in name
    and self.config.tie_word_embeddings  # <-- added check
):

[gemini-code-assist]

Warning

You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again!

[JustinTong0323]

It seems not a valid fix, my test still shows garbage output.

[JustinTong0323]

/tag-and-rerun-ci

[Lollipop]

@JustinTong0323 Thanks for testing! The previous fix was incorrect - I've updated the PR with the correct fix now.

Root Cause Clarification

The issue only affects models with tie_word_embeddings=False. Here's the difference:

Model	tie_word_embeddings	Affected
Qwen3-VL-2B	True	No
Qwen3-VL-4B	True	No
Qwen3-VL-8B	False	Yes

For models with tie_word_embeddings=True, embed_tokens and lm_head share the same weights, so copying is correct.

For Qwen3-VL-8B (tie_word_embeddings=False), lm_head has its own independent weights. The unconditional copy overwrites these weights with embed_tokens, causing garbage output.

Updated Fix

The new fix adds a check for self.config.tie_word_embeddings:

if (
    self.pp_group.is_last_rank
    and "model.embed_tokens.weight" in name
    and self.config.tie_word_embeddings  # <-- only copy when weights are shared
):

Could you please test again with Qwen3-VL-8B specifically? The 2B/4B models should work fine with or without this fix.

[JustinTong0323]

LGTM