Was two phase decoding actually used for Qwen3-8B?

[cheongalc]

Dear authors, thank you for releasing TTT-Discover.

I am reimplementing the method and ran into some ambiguity around the generation path for Qwen3-8B.

From reading the released code, I see that the rollout code appears to instantiate the same TwoPhaseTokenCompleter regardless of model choice:

discover/ttt_discover/rl/train.py

Lines 340 to 345 in 5df1a0e

	policy = TwoPhaseTokenCompleter(
	sampling_client=sampling_client,
	tokenizer=tokenizer,
	phase1_max_tokens=phase1_max_tokens,
	temperature=temperature,
	)

However, that completer appears to be written specifically for GPT-OSS:

discover/ttt_discover/tinker_utils/completers.py

Lines 60 to 64 in 5df1a0e

	PHASE2_PREFILL = "\n\n... okay, I am out of thinking tokens. I need to send my final message now."
	# Full marker to transition from analysis to final channel
	GPTOSS_FINAL_MARKER = "<|end|><|start|>assistant<|channel|>final<|message|>"
	# Marker that indicates we're already in the final channel
	GPTOSS_FINAL_CHANNEL_INDICATOR = "<|channel|>final<|message|>"

discover/ttt_discover/tinker_utils/completers.py

Line 120 in 5df1a0e

if self._contains_subsequence(phase1_tokens, self.GPTOSS_FINAL_CHANNEL_INDICATOR):

Meanwhile, the Qwen3 renderer uses a different chat format, and stops on <|im_end|>. Thus, I'm not sure whether the two phase decoding is silently skipped when Qwen3-8B is used for the math and circle packing tasks.

Could I check if the intended behavior is any of the following:

• there is a true two-phase decoding with a Qwen-specific string to do the token forcing at token 26,000, or
• the two-phase token completer runs as is, but this means it tries to match GPT-OSS specific harmony channel tags to a string that follows the Qwen chat template, so those conditions silently fail, or
• Qwen3-8B just uses ordinary single-phase decoding.

Thank you!

[mertyg]

thanks for the quick check! that's right, as it is, "Qwen3-8B just uses ordinary single-phase decoding". we'd welcome contributions if you'd like to make 2-phase work for qwen!