bugfix: Fix prompt/context length limits and NPU group initialization.

[sanlio36]

No description provided.

[gemini-code-assist]

Code Review

This pull request refactors prompt length validation across the LLMMaster, RecMaster, and VLMMaster classes, and ensures max_tokens_per_batch is correctly propagated in the Master configuration. Additionally, it updates CollectiveCommunicator to support specific host resolution for NPU-based MoE tensor parallel groups. The review identifies recurring issues in the new validation logic, specifically signed/unsigned comparison violations of the style guide (Rule 79) and off-by-one errors in the prompt length checks that could lead to incorrect request rejection or insufficient space for token generation.

[gemini-code-assist]

The prompt length check has two issues:

1. Signed/Unsigned Comparison: Comparing local_prompt_tokens.size() (unsigned) with prompt_token_limit (signed) violates the repository style guide (Rule 79), which requires explicit static_cast for type conversions.
2. Off-by-one Error: When enable_chunked_prefill is false, the prompt must fit within max_tokens_per_batch. A prompt of size exactly equal to the batch size is valid for prefill, but the >= check incorrectly rejects it. Conversely, for the model's max_position_embeddings, the check should be strict (size < max_context_len) because there must be room for at least one generated token.

Adjusting the limit and using a > comparison correctly handles both constraints.

Suggested change

	int32_t prompt_token_limit = max_context_len;
	if (!options_.enable_chunked_prefill()) {
	max_context_len =
	std::min(max_context_len, options_.max_tokens_per_batch());
	prompt_token_limit =
	std::min(prompt_token_limit, options_.max_tokens_per_batch());
	}
	if (local_prompt_tokens.size() >= max_context_len) {
	if (local_prompt_tokens.size() >= prompt_token_limit) {
	int32_t prompt_token_limit = max_context_len - 1;
	if (!options_.enable_chunked_prefill()) {
	prompt_token_limit =
	std::min(prompt_token_limit, options_.max_tokens_per_batch());
	}
	if (static_cast<int32_t>(local_prompt_tokens.size()) > prompt_token_limit) {

References

1. Use static_cast for all type conversions. Never use C-style casts. ^(link)

[gemini-code-assist]

The prompt length check has two issues:

1. Signed/Unsigned Comparison: Comparing prompt_tokens.size() (unsigned) with prompt_token_limit (signed) violates the repository style guide (Rule 79).
2. Off-by-one Error: The >= check incorrectly rejects prompts that exactly fill the batch/request limit when chunked prefill is disabled. For the model limit, the check should be strict (size < max_context_len), but for the batch limit, it should be inclusive (size <= max_tokens_per_req).

Suggested change

	int32_t prompt_token_limit = max_context_len;
	if (!options_.enable_chunked_prefill()) {
	int32_t max_tokens_per_req = options_.max_tokens_per_batch();
	if (rec_type_ == RecType::kLlmRec && is_rec_multi_round_mode()) {
	CHECK_GT(options_.max_seqs_per_batch(), 0)
	<< "max_seqs_per_batch must be greater than 0 in multi-round mode";
	max_tokens_per_req /= options_.max_seqs_per_batch();
	}
	max_context_len = std::min(max_context_len, max_tokens_per_req);
	prompt_token_limit = std::min(prompt_token_limit, max_tokens_per_req);
	}
	if (prompt_tokens.size() >= max_context_len) {
	if (prompt_tokens.size() >= prompt_token_limit) {
	int32_t prompt_token_limit = max_context_len - 1;
	if (!options_.enable_chunked_prefill()) {
	int32_t max_tokens_per_req = options_.max_tokens_per_batch();
	if (rec_type_ == RecType::kLlmRec && is_rec_multi_round_mode()) {
	CHECK_GT(options_.max_seqs_per_batch(), 0)
	<< "max_seqs_per_batch must be greater than 0 in multi-round mode";
	max_tokens_per_req /= options_.max_seqs_per_batch();
	}
	prompt_token_limit = std::min(prompt_token_limit, max_tokens_per_req);
	}
	if (static_cast<int32_t>(prompt_tokens.size()) > prompt_token_limit) {

References

1. Use static_cast for all type conversions. Never use C-style casts. ^(link)

[gemini-code-assist]

The prompt length check has two issues:

1. Signed/Unsigned Comparison: Comparing prompt_tokens.size() (unsigned) with prompt_token_limit (signed) violates the repository style guide (Rule 79).
2. Off-by-one Error: The >= check incorrectly rejects prompts that exactly fill the batch limit when chunked prefill is disabled. Adjusting the limit and using a > comparison correctly handles both the model context limit and the batch size limit.

Suggested change

	int32_t prompt_token_limit = max_context_len;
	if (!options_.enable_chunked_prefill()) {
	max_context_len =
	std::min(max_context_len, options_.max_tokens_per_batch());
	prompt_token_limit =
	std::min(prompt_token_limit, options_.max_tokens_per_batch());
	}
	if (prompt_tokens.size() >= max_context_len) {
	if (prompt_tokens.size() >= prompt_token_limit) {
	int32_t prompt_token_limit = max_context_len - 1;
	if (!options_.enable_chunked_prefill()) {
	prompt_token_limit =
	std::min(prompt_token_limit, options_.max_tokens_per_batch());
	}
	if (static_cast<int32_t>(prompt_tokens.size()) > prompt_token_limit) {

References

1. Use static_cast for all type conversions. Never use C-style casts. ^(link)