draft: enable LM tp for MTP, under attention dp case #7344
Conversation
Signed-off-by: Dongxu Yang <[email protected]>
📝 WalkthroughTip 🔌 Remote MCP (Model Context Protocol) integration is now available!Pro plan users can now connect to remote MCP servers from the Integrations page. Connect with popular remote MCPs such as Notion and Linear to add more context to your reviews and chats. ✨ Finishing Touches
🧪 Generate unit tests
Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out. 🪧 TipsChatThere are 3 ways to chat with CodeRabbit:
SupportNeed help? Create a ticket on our support page for assistance with any issues or questions. CodeRabbit Commands (Invoked using PR/Issue comments)Type Other keywords and placeholders
Status, Documentation and Community
|
![coderabbitai[bot]](https://avatars.githubusercontent.com/in/347564?s=60&v=4){kind=small}
There was a problem hiding this comment.
Actionable comments posted: 6
Caution
Some comments are outside the diff and can’t be posted inline due to platform limitations.
⚠️ Outside diff range comments (1)
tensorrt_llm/_torch/models/modeling_utils.py (1)
398-406: Ensure embed_tokens is created with the samemapping/TP settings aslm_headunder ADP+LM-TP whentie_word_embeddings=True.
Currently most models doif config.mapping.is_first_pp_rank(): self.vocab_embedding = Embedding(…, dtype=config.dtype)omitting
mapping=config.mapping(andtensor_parallel_mode=…), soembed_tokens.tp_size==1even whenlm_head.tp_size>1, triggering the asserts in modeling_utils.py. Pass the samemapping(and TP mode) into everyEmbedding(…)call for vocab/embed_tokens under ADP+LM-TP, or disabletie_word_embeddingswhen TP settings differ.
🧹 Nitpick comments (8)
tensorrt_llm/_torch/models/modeling_speculative.py (1)
436-440: Remove noisy debug prints; gate with a debug flag if needed.Printing per-step in distributed inference is costly and pollutes logs. Prefer logger with a runtime flag (e.g., enable_llm_debug()) or remove entirely.
Apply:
- # print(f"lm_head.weight.data_ptr: {self.lm_head.weight.data_ptr()}") - # print(f"lm_head.weight.shape: {self.lm_head.weight.shape}") - print(f"In SpecDecOneEngineForCausalLM, before spec_worker, logits.shape: {logits.shape}") - # print(f"draft_model.lm_head.weight.data_ptr: {self.draft_model.lm_head.weight.data_ptr()}") - # print(f"draft_model.lm_head.weight.shape: {self.draft_model.lm_head.weight.shape}") + # Debug: add gated logging here if absolutely necessary.tensorrt_llm/_torch/models/modeling_deepseekv3.py (2)
167-167: Remove redundant import inside hot path.allgather is already imported at module scope; avoid re-import in forward.
Apply:
- from ..distributed import allgather
1-27: Add NVIDIA 2025 SPDX header per repo guidelines.This file lacks the required NVIDIA SPDX header. Keep third‑party license block, but prepend the NVIDIA header.
Apply at file top:
+# SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# SPDX-License-Identifier: Apache-2.0tensorrt_llm/_torch/modules/embedding.py (2)
72-74: Remove unconditional prints; they sync/slow and violate lint (E501).Use a logger gated by a runtime flag, or drop them.
Apply:
- print('*' * 50) - print(f"enable_attention_dp: {mapping.enable_attention_dp}, enable_lm_tp_in_adp: {getattr(mapping, 'enable_lm_tp_in_adp', False)}") - print(f"In LMHead, weight_shape: {weight_shape}") + # Optional: add gated debug logging here if necessary.
1-1: Add NVIDIA 2025 SPDX header per project policy.File lacks required header.
Apply at file top:
+# SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# SPDX-License-Identifier: Apache-2.0tensorrt_llm/_torch/models/modeling_utils.py (2)
355-364: LMHead ADP/LM‑TP gating looks good; drop prints.Creation paths match the new flag. Replace prints with logger or remove.
Apply:
- if config.mapping.enable_attention_dp and not getattr(config.mapping, 'enable_lm_tp_in_adp', False): - print(f"In DecoderModelForCausalLM, creating LMHead without TP") + if config.mapping.enable_attention_dp and not getattr(config.mapping, 'enable_lm_tp_in_adp', False): self.lm_head = LMHead( @@ - else: - print(f"In DecoderModelForCausalLM, creating LMHead with TP") + else: # TODO(zhenhuanc): Currently lm_head Linear will not accept QuantConfig
1-1: Add NVIDIA 2025 SPDX header per policy.Header missing.
Apply at file top:
+# SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# SPDX-License-Identifier: Apache-2.0tensorrt_llm/_torch/speculative/mtp.py (1)
1122-1143: ADP + LM TP handling looks correct but needs optimization.The new branch for ADP + LM TP correctly:
- Gathers logits across all ranks
- Reshapes based on TP configuration
- Slices by current rank
- Computes argmax on the sliced data
However, the approach could be optimized to reduce memory usage.
Consider using the same optimized approach as the non-ADP branch (lines 1115-1121) that combines local max values with indices before gathering, rather than gathering all logits which is more memory intensive.
📜 Review details
Configuration used: Path: .coderabbit.yaml
Review profile: CHILL
Plan: Pro
💡 Knowledge Base configuration:
- MCP integration is disabled by default for public repositories
- Jira integration is disabled by default for public repositories
- Linear integration is disabled by default for public repositories
You can enable these sources in your CodeRabbit configuration.
📒 Files selected for processing (8)
tensorrt_llm/_torch/models/modeling_deepseekv3.py(1 hunks)tensorrt_llm/_torch/models/modeling_speculative.py(1 hunks)tensorrt_llm/_torch/models/modeling_utils.py(2 hunks)tensorrt_llm/_torch/modules/embedding.py(2 hunks)tensorrt_llm/_torch/modules/logits_processor.py(2 hunks)tensorrt_llm/_torch/speculative/mtp.py(6 hunks)tensorrt_llm/llmapi/llm_args.py(4 hunks)tensorrt_llm/mapping.py(2 hunks)
🧰 Additional context used
📓 Path-based instructions (2)
**/*.py
📄 CodeRabbit inference engine (CODING_GUIDELINES.md)
**/*.py: Code must target Python 3.8+
Indent with 4 spaces; do not use tabs
Preserve module namespaces in imports: import the subpackage/module, not the symbol (from package.subpackage import foo; foo.SomeClass())
Naming: files snake_case; classes PascalCase; functions/methods snake_case; local variables snake_case (k_ prefix if starting with a number); globals G_ + UPPER_SNAKE_CASE; constants UPPER_SNAKE_CASE
Avoid shadowing outer-scope variables; initialize all externally visible members in init
Prefer docstrings for interfaces used outside a file; reserve comments for function-internal or file-local interfaces
Use Google-style docstrings for classes and functions; inline docstrings for attributes/variables are allowed
Avoid reflection when straightforward code suffices (e.g., prefer explicit parameters over dict(**locals()))
Use narrow except clauses (e.g., catch FileNotFoundError instead of bare except)
For duck-typing try/except, keep try body minimal and use else for the main logic
Files:
tensorrt_llm/_torch/models/modeling_speculative.pytensorrt_llm/_torch/modules/embedding.pytensorrt_llm/mapping.pytensorrt_llm/llmapi/llm_args.pytensorrt_llm/_torch/models/modeling_utils.pytensorrt_llm/_torch/modules/logits_processor.pytensorrt_llm/_torch/models/modeling_deepseekv3.pytensorrt_llm/_torch/speculative/mtp.py
**/*.{cpp,cc,cxx,cu,h,hpp,hh,hxx,cuh,py}
📄 CodeRabbit inference engine (CODING_GUIDELINES.md)
Prepend NVIDIA copyright header with current year to all source files
Files:
tensorrt_llm/_torch/models/modeling_speculative.pytensorrt_llm/_torch/modules/embedding.pytensorrt_llm/mapping.pytensorrt_llm/llmapi/llm_args.pytensorrt_llm/_torch/models/modeling_utils.pytensorrt_llm/_torch/modules/logits_processor.pytensorrt_llm/_torch/models/modeling_deepseekv3.pytensorrt_llm/_torch/speculative/mtp.py
🧠 Learnings (1)
📚 Learning: 2025-08-14T06:36:40.701Z
Learnt from: timlee0212
PR: NVIDIA/TensorRT-LLM#6886
File: tensorrt_llm/_torch/models/modeling_deepseekv3.py:0-0
Timestamp: 2025-08-14T06:36:40.701Z
Learning: In DeepSeek V3 model (tensorrt_llm/_torch/models/modeling_deepseekv3.py), the disagreement between AllReduce.__init__ guard and _compute_mlp_tp_size logic for MNNVL usage is expected by design. The AllReduce component and MLP TP-size computation intentionally use different criteria for MNNVL availability decisions.
Applied to files:
tensorrt_llm/_torch/models/modeling_deepseekv3.py
🧬 Code graph analysis (4)
tensorrt_llm/_torch/models/modeling_utils.py (2)
tensorrt_llm/_torch/modules/embedding.py (1)
LMHead(15-118)tensorrt_llm/_torch/modules/logits_processor.py (1)
LogitsProcessor(10-75)
tensorrt_llm/_torch/modules/logits_processor.py (4)
tensorrt_llm/_torch/attention_backend/interface.py (1)
AttentionMetadata(39-328)tensorrt_llm/_torch/modules/linear.py (1)
Linear(1495-1704)tensorrt_llm/_torch/distributed/communicator.py (1)
tp_size(46-47)tensorrt_llm/mapping.py (1)
tp_rank(340-341)
tensorrt_llm/_torch/models/modeling_deepseekv3.py (2)
tensorrt_llm/_torch/distributed/ops.py (1)
allgather(141-225)tensorrt_llm/functional.py (1)
allgather(4142-4226)
tensorrt_llm/_torch/speculative/mtp.py (4)
tensorrt_llm/_torch/distributed/communicator.py (4)
tp_size(46-47)allgather(94-95)allgather(107-108)tp_rank(54-55)tensorrt_llm/_torch/distributed/ops.py (1)
allgather(141-225)cpp/tensorrt_llm/thop/allgatherOp.cpp (2)
allgather(122-137)allgather(122-122)tensorrt_llm/mapping.py (1)
tp_rank(340-341)
🪛 Ruff (0.12.2)
tensorrt_llm/_torch/modules/embedding.py
73-73: Line too long (139 > 120)
(E501)
tensorrt_llm/_torch/speculative/mtp.py
865-865: Line too long (134 > 120)
(E501)
868-868: Line too long (158 > 120)
(E501)
1139-1139: Line too long (140 > 120)
(E501)
🔇 Additional comments (13)
tensorrt_llm/_torch/models/modeling_deepseekv3.py (1)
171-177: Confirm downstream handles sharded logits correctly.You temporarily set lm_head.gather_output=False, compute logits, then re-enable it, but the returned logits here are still shard-local. Ensure a later stage all-gathers logits along vocab or keep gather_output=True here.
Would you like me to scan for consumers of DeepseekV3MTPHead.forward to verify a subsequent vocab-dim all-gather?
tensorrt_llm/_torch/modules/embedding.py (1)
40-42: Gating LM TP under ADP is correct.The guard allows LM TP only when explicitly enabled. Good defensive default.
tensorrt_llm/_torch/models/modeling_utils.py (1)
407-407: Init LogitsProcessor with config: OK.Constructor change is reflected here; no issues.
tensorrt_llm/llmapi/llm_args.py (3)
228-228: LGTM! The new field enables LM tensor parallelism in attention data parallel mode.The addition of
enable_lm_tp_in_adpaligns with the PR objectives to enable LM TP for MTP under attention DP case.
1251-1255: Good addition for enabling LM TP in attention DP mode.The field is correctly marked with
status="beta"to indicate it's not yet stable, and the description is clear.
292-292: Proper propagation of the new flag to Mapping.The
enable_lm_tp_in_adpflag is correctly propagated from the internal_ParallelConfigto the publicMappingobject, ensuring downstream components can access this configuration.Also applies to: 1503-1503
tensorrt_llm/_torch/modules/logits_processor.py (5)
12-14: Constructor properly accepts ModelConfig for distributed logic paths.The addition of
model_configparameter enables the LogitsProcessor to access mapping configuration for ADP LM-TP scenarios.
33-34: Token count calculation looks correct.The computation properly handles the reshaping to get the batch dimension.
36-52: All-gather logic for ADP + LM TP looks correct.The implementation correctly:
- Checks for both
enable_attention_dpandenable_lm_tp_in_adpflags- Pads hidden states to match max tokens across ranks
- Performs all-gather operation along the token dimension
Good use of
getattrwith a default value for backward compatibility.
55-63: Clean toggling of gather_output flag.The logic correctly disables
gather_outputduring LM head forward when appropriate, preventing double gathering.
67-72: Tensor manipulation for TP looks correct.The reshaping and slicing logic properly handles the distributed tensor layout.
tensorrt_llm/_torch/speculative/mtp.py (2)
477-493: Proper padding of hidden states for uniform token count.The padding logic correctly aligns hidden states to
all_rank_max_num_tokensbefore passing to the shared head, ensuring consistent tensor dimensions across ranks.
1246-1246: Properly passes iter parameter to draft_sampler.The change correctly passes the iteration index to the updated
draft_samplermethod.
| # Add pre-lm gather logic | ||
| if (self.model_config.mapping.enable_attention_dp and | ||
| getattr(self.model_config.mapping, 'enable_lm_tp_in_adp', False)): | ||
| # ADP + LM TP mode: perform All-Gather before LM_head | ||
| from ..distributed import allgather | ||
| hidden_states = allgather(hidden_states, self.model_config.mapping, dim=0) | ||
|
|
There was a problem hiding this comment.
Fix potential shape mismatch in ADP all-gather by passing per-rank sizes.
Without sizes, allgather requires identical lengths across ranks. In ADP, per-rank token counts can differ; this risks incorrect results or runtime assertions.
Apply:
- # Add pre-lm gather logic
- if (self.model_config.mapping.enable_attention_dp and
- getattr(self.model_config.mapping, 'enable_lm_tp_in_adp', False)):
- # ADP + LM TP mode: perform All-Gather before LM_head
- from ..distributed import allgather
- hidden_states = allgather(hidden_states, self.model_config.mapping, dim=0)
+ # Add pre-lm gather logic
+ if (self.model_config.mapping.enable_attention_dp and
+ getattr(self.model_config.mapping, 'enable_lm_tp_in_adp', False)):
+ # ADP + LM TP mode: perform All-Gather before LM_head with per-rank sizes
+ hidden_states = allgather(
+ hidden_states,
+ self.model_config.mapping,
+ dim=0,
+ sizes=getattr(attn_metadata, "all_rank_num_tokens", None),
+ )Committable suggestion skipped: line range outside the PR's diff.
| logits = logits.view(self.model_config.mapping.tp_size, | ||
| local_batch_size, -1) | ||
| logits = logits[self.model_config.mapping.tp_rank][:token_count] | ||
| print(f"In LogitsProcessor, final logits.shape: {logits.shape}") |
There was a problem hiding this comment.
Remove or guard debug print statement.
Debug print statements should not be left in production code.
- print(f"In LogitsProcessor, final logits.shape: {logits.shape}")
+ # Uncomment for debugging:
+ # print(f"In LogitsProcessor, final logits.shape: {logits.shape}")📝 Committable suggestion
‼️ IMPORTANT
Carefully review the code before committing. Ensure that it accurately replaces the highlighted code, contains no missing lines, and has no issues with indentation. Thoroughly test & benchmark the code to ensure it meets the requirements.
| print(f"In LogitsProcessor, final logits.shape: {logits.shape}") | |
| # Uncomment for debugging: | |
| # print(f"In LogitsProcessor, final logits.shape: {logits.shape}") |
🤖 Prompt for AI Agents
In tensorrt_llm/_torch/modules/logits_processor.py around line 73, remove the
stray debug print statement or replace it with a guarded debug-level log: either
delete the print entirely or call the module/logger debug method (e.g.,
logger.debug) or wrap it behind a runtime debug/config flag so that it only runs
in development; ensure no direct stdout prints remain in production code.
| print( | ||
| f"In sample_and_accept_draft_tokens, target_tokens.shape: {target_tokens.shape}" | ||
| ) | ||
|
|
||
| # context | ||
| accepted_tokens[:num_contexts, 0] = target_tokens[:num_contexts] | ||
|
|
||
| print( | ||
| f"In sample_and_accept_draft_tokens, accepted_tokens.shape: {accepted_tokens.shape}, num_contexts: {num_contexts}" | ||
| ) | ||
| print( | ||
| f"In sample_and_accept_draft_tokens, target_tokens.shape: {target_tokens.shape}, num_gens: {num_gens}, mtp_num_modules: {mtp_num_modules}" | ||
| ) |
There was a problem hiding this comment.
Remove debug print statements from production code.
Multiple debug print statements should be removed or properly guarded.
- print(
- f"In sample_and_accept_draft_tokens, target_tokens.shape: {target_tokens.shape}"
- )
+ # Uncomment for debugging:
+ # print(f"In sample_and_accept_draft_tokens, target_tokens.shape: {target_tokens.shape}")
# context
accepted_tokens[:num_contexts, 0] = target_tokens[:num_contexts]
- print(
- f"In sample_and_accept_draft_tokens, accepted_tokens.shape: {accepted_tokens.shape}, num_contexts: {num_contexts}"
- )
- print(
- f"In sample_and_accept_draft_tokens, target_tokens.shape: {target_tokens.shape}, num_gens: {num_gens}, mtp_num_modules: {mtp_num_modules}"
- )
+ # Uncomment for debugging:
+ # print(f"In sample_and_accept_draft_tokens, accepted_tokens.shape: {accepted_tokens.shape}, num_contexts: {num_contexts}")
+ # print(f"In sample_and_accept_draft_tokens, target_tokens.shape: {target_tokens.shape}, num_gens: {num_gens}, mtp_num_modules: {mtp_num_modules}")📝 Committable suggestion
‼️ IMPORTANT
Carefully review the code before committing. Ensure that it accurately replaces the highlighted code, contains no missing lines, and has no issues with indentation. Thoroughly test & benchmark the code to ensure it meets the requirements.
| print( | |
| f"In sample_and_accept_draft_tokens, target_tokens.shape: {target_tokens.shape}" | |
| ) | |
| # context | |
| accepted_tokens[:num_contexts, 0] = target_tokens[:num_contexts] | |
| print( | |
| f"In sample_and_accept_draft_tokens, accepted_tokens.shape: {accepted_tokens.shape}, num_contexts: {num_contexts}" | |
| ) | |
| print( | |
| f"In sample_and_accept_draft_tokens, target_tokens.shape: {target_tokens.shape}, num_gens: {num_gens}, mtp_num_modules: {mtp_num_modules}" | |
| ) | |
| # Uncomment for debugging: | |
| # print(f"In sample_and_accept_draft_tokens, target_tokens.shape: {target_tokens.shape}") | |
| # context | |
| accepted_tokens[:num_contexts, 0] = target_tokens[:num_contexts] | |
| # Uncomment for debugging: | |
| # print(f"In sample_and_accept_draft_tokens, accepted_tokens.shape: {accepted_tokens.shape}, num_contexts: {num_contexts}") | |
| # print(f"In sample_and_accept_draft_tokens, target_tokens.shape: {target_tokens.shape}, num_gens: {num_gens}, mtp_num_modules: {mtp_num_modules}") |
🧰 Tools
🪛 Ruff (0.12.2)
865-865: Line too long (134 > 120)
(E501)
868-868: Line too long (158 > 120)
(E501)
🤖 Prompt for AI Agents
In tensorrt_llm/_torch/speculative/mtp.py around lines 858 to 869, there are
multiple raw print() debug statements; remove them or replace them with a proper
logger.debug() (or guard them behind an existing debug flag) so production code
does not emit uncontrolled stdout. Specifically, delete the three print(...)
calls or change them to process-wide logger.debug(...) with contextual messages
including variable values, or wrap them in if self.debug / if debug_mode: ... so
they only run in debug mode; ensure any used logger is already configured and
imported.
| @@ -1074,6 +1097,7 @@ def get_draft_tokens_from_gathered(self, gathered): | |||
| def draft_sampler( | |||
| self, | |||
| logits: torch.Tensor, | |||
| iter: int, | |||
There was a problem hiding this comment.
💡 Verification agent
🧩 Analysis chain
Breaking change: draft_sampler now requires iter parameter.
The addition of the iter parameter changes the public API of draft_sampler. Ensure all callers are updated accordingly.
🏁 Script executed:
#!/bin/bash
# Search for all calls to draft_sampler to ensure they pass the iter parameter
rg -n "draft_sampler\(" --type pyLength of output: 306
Include the new iter argument in all draft_sampler calls
tensorrt_llm/_torch/speculative/mtp.py:492 — update the call from
self.draft_sampler(logits)to
self.draft_sampler(logits, i)🤖 Prompt for AI Agents
tensorrt_llm/_torch/speculative/mtp.py around lines 492 and anywhere else where
draft_sampler is invoked: the new function signature includes a second parameter
iter: int, so update all calls like self.draft_sampler(logits) to pass the
current loop/index variable (e.g., self.draft_sampler(logits, i) or
self.draft_sampler(logits, iter) depending on your local variable name). Ensure
the variable you pass exists in scope and matches the intended iteration
counter, and update any tests/call sites accordingly.
| elif (self.model_config is not None | ||
| and hasattr(self.model_config, 'mapping') | ||
| and self.model_config.mapping.tp_size > 1) and ( | ||
| self.model_config.mapping.enable_attention_dp and getattr( | ||
| self.model_config.mapping, 'enable_lm_tp_in_adp', False)): | ||
| # For ADP + LM TP mode, we need to find the global argmax across all TP ranks | ||
| # First, get local argmax and max values | ||
| # print(f"In draft_sampler, initial logits.shape: {logits.shape}") | ||
| # combined = self.get_local_max_and_combined(logits) | ||
| gathered = allgather(logits, self.model_config.mapping, dim=-1) | ||
| batch_size = logits.shape[0] | ||
| local_batch_size = batch_size // self.model_config.mapping.tp_size | ||
| gathered = gathered.view(self.model_config.mapping.tp_size, | ||
| local_batch_size, -1) | ||
| sliced_gathered = gathered[self.model_config.mapping.tp_rank] | ||
| # print(f"In draft_sampler, gathered.shape: {gathered.shape}") | ||
| print( | ||
| f"In draft_sampler, iter: {iter}, rank: {self.model_config.mapping.tp_rank}, sliced_gathered.shape: {sliced_gathered.shape}" | ||
| ) | ||
| # draft_tokens = self.get_draft_tokens_from_gathered(sliced_gathered) | ||
| draft_tokens = torch.argmax(sliced_gathered, | ||
| dim=-1).type(torch.int32) |
There was a problem hiding this comment.
Remove debug print statement.
Line 1139 contains a debug print that should be removed.
- print(
- f"In draft_sampler, iter: {iter}, rank: {self.model_config.mapping.tp_rank}, sliced_gathered.shape: {sliced_gathered.shape}"
- )
+ # Uncomment for debugging:
+ # print(f"In draft_sampler, iter: {iter}, rank: {self.model_config.mapping.tp_rank}, sliced_gathered.shape: {sliced_gathered.shape}")📝 Committable suggestion
‼️ IMPORTANT
Carefully review the code before committing. Ensure that it accurately replaces the highlighted code, contains no missing lines, and has no issues with indentation. Thoroughly test & benchmark the code to ensure it meets the requirements.
| elif (self.model_config is not None | |
| and hasattr(self.model_config, 'mapping') | |
| and self.model_config.mapping.tp_size > 1) and ( | |
| self.model_config.mapping.enable_attention_dp and getattr( | |
| self.model_config.mapping, 'enable_lm_tp_in_adp', False)): | |
| # For ADP + LM TP mode, we need to find the global argmax across all TP ranks | |
| # First, get local argmax and max values | |
| # print(f"In draft_sampler, initial logits.shape: {logits.shape}") | |
| # combined = self.get_local_max_and_combined(logits) | |
| gathered = allgather(logits, self.model_config.mapping, dim=-1) | |
| batch_size = logits.shape[0] | |
| local_batch_size = batch_size // self.model_config.mapping.tp_size | |
| gathered = gathered.view(self.model_config.mapping.tp_size, | |
| local_batch_size, -1) | |
| sliced_gathered = gathered[self.model_config.mapping.tp_rank] | |
| # print(f"In draft_sampler, gathered.shape: {gathered.shape}") | |
| print( | |
| f"In draft_sampler, iter: {iter}, rank: {self.model_config.mapping.tp_rank}, sliced_gathered.shape: {sliced_gathered.shape}" | |
| ) | |
| # draft_tokens = self.get_draft_tokens_from_gathered(sliced_gathered) | |
| draft_tokens = torch.argmax(sliced_gathered, | |
| dim=-1).type(torch.int32) | |
| elif (self.model_config is not None | |
| and hasattr(self.model_config, 'mapping') | |
| and self.model_config.mapping.tp_size > 1) and ( | |
| self.model_config.mapping.enable_attention_dp and getattr( | |
| self.model_config.mapping, 'enable_lm_tp_in_adp', False)): | |
| # For ADP + LM TP mode, we need to find the global argmax across all TP ranks | |
| # First, get local argmax and max values | |
| # print(f"In draft_sampler, initial logits.shape: {logits.shape}") | |
| # combined = self.get_local_max_and_combined(logits) | |
| gathered = allgather(logits, self.model_config.mapping, dim=-1) | |
| batch_size = logits.shape[0] | |
| local_batch_size = batch_size // self.model_config.mapping.tp_size | |
| gathered = gathered.view(self.model_config.mapping.tp_size, | |
| local_batch_size, -1) | |
| sliced_gathered = gathered[self.model_config.mapping.tp_rank] | |
| # print(f"In draft_sampler, gathered.shape: {gathered.shape}") | |
| # Uncomment for debugging: | |
| # print(f"In draft_sampler, iter: {iter}, rank: {self.model_config.mapping.tp_rank}, sliced_gathered.shape: {sliced_gathered.shape}") | |
| # draft_tokens = self.get_draft_tokens_from_gathered(sliced_gathered) | |
| draft_tokens = torch.argmax(sliced_gathered, | |
| dim=-1).type(torch.int32) |
🧰 Tools
🪛 Ruff (0.12.2)
1139-1139: Line too long (140 > 120)
(E501)
🤖 Prompt for AI Agents
In tensorrt_llm/_torch/speculative/mtp.py around lines 1122 to 1143, remove the
debug print statement at line 1139 (the print starting with "In draft_sampler,
iter: {iter}...") — delete that print call and, if needed for future debugging,
replace it with a proper logger.debug call using the module/logger used
elsewhere; otherwise simply remove the line so no raw print remains.
| enable_lm_tp_in_adp=False): | ||
| # set default values for non-moe cases |
There was a problem hiding this comment.
Include new flag in equality/hash/serialization; otherwise caches/config round-trips break.
enable_lm_tp_in_adp is stored but omitted from eq, hash, and to_dict. Two Mapping instances with different values will compare equal and serialize without the flag.
Apply:
@@
def __eq__(self, other):
@@
- return (self.world_size == other.world_size and self.rank == other.rank
+ return (self.world_size == other.world_size and self.rank == other.rank
and self.gpus_per_node == other.gpus_per_node
and self.cp_size == other.cp_size
and self.tp_size == other.tp_size
and self.moe_cluster_size == other.moe_cluster_size
and self.pp_size == other.pp_size
and self.moe_tp_size == other.moe_tp_size
and self.moe_ep_size == other.moe_ep_size
and self.attn_tp_size == other.attn_tp_size
and self.attn_cp_size == other.attn_cp_size
and self.cp_config == other.cp_config
- and self.auto_parallel == other.auto_parallel)
+ and self.auto_parallel == other.auto_parallel
+ and self.enable_attention_dp == other.enable_attention_dp
+ and self.enable_lm_tp_in_adp == other.enable_lm_tp_in_adp)
@@
def __hash__(self):
return hash((
self.world_size,
self.rank,
self.gpus_per_node,
self.cp_size,
self.tp_size,
self.pp_size,
self.moe_tp_size,
self.moe_cluster_size,
self.moe_ep_size,
self.attn_tp_size,
self.attn_cp_size,
# note: we do not allow updating cp_config after initialization
tuple(sorted(self.cp_config.items())),
self.auto_parallel,
+ self.enable_attention_dp,
+ self.enable_lm_tp_in_adp,
))
@@
def to_dict(self):
return {
'world_size': self.world_size,
'rank': self.rank,
'gpus_per_node': self.gpus_per_node,
'cp_size': self.cp_size,
'tp_size': self.tp_size,
'pp_size': self.pp_size,
'moe_tp_size': self.moe_tp_size,
'moe_cluster_size': self.moe_cluster_size,
'moe_ep_size': self.moe_ep_size,
'attn_tp_size': self.attn_tp_size,
'attn_cp_size': self.attn_cp_size,
'cp_config': self.cp_config,
'auto_parallel': self.auto_parallel,
+ 'enable_attention_dp': self.enable_attention_dp,
+ 'enable_lm_tp_in_adp': self.enable_lm_tp_in_adp,
}Also applies to: 227-229
Summary by CodeRabbit
New Features
Chores
Public API
Description
Test Coverage
PR Checklist
Please review the following before submitting your PR:
PR description clearly explains what and why. If using CodeRabbit's summary, please make sure it makes sense.
PR Follows TRT-LLM CODING GUIDELINES to the best of your knowledge.
Test cases are provided for new code paths (see test instructions)
Any new dependencies have been scanned for license and vulnerabilities
CODEOWNERS updated if ownership changes
Documentation updated as needed
The reviewers assigned automatically/manually are appropriate for the PR.
Please check this after reviewing the above items as appropriate for this PR.
GitHub Bot Help
/bot [-h] ['run', 'kill', 'skip', 'reuse-pipeline'] ...Provide a user friendly way for developers to interact with a Jenkins server.
Run
/bot [-h|--help]to print this help message.See details below for each supported subcommand.
run [--reuse-test (optional)pipeline-id --disable-fail-fast --skip-test --stage-list "A10-PyTorch-1, xxx" --gpu-type "A30, H100_PCIe" --test-backend "pytorch, cpp" --add-multi-gpu-test --only-multi-gpu-test --disable-multi-gpu-test --post-merge --extra-stage "H100_PCIe-TensorRT-Post-Merge-1, xxx" --detailed-log --debug(experimental)]Launch build/test pipelines. All previously running jobs will be killed.
--reuse-test (optional)pipeline-id(OPTIONAL) : Allow the new pipeline to reuse build artifacts and skip successful test stages from a specified pipeline or the last pipeline if no pipeline-id is indicated. If the Git commit ID has changed, this option will be always ignored. The DEFAULT behavior of the bot is to reuse build artifacts and successful test results from the last pipeline.--disable-reuse-test(OPTIONAL) : Explicitly prevent the pipeline from reusing build artifacts and skipping successful test stages from a previous pipeline. Ensure that all builds and tests are run regardless of previous successes.--disable-fail-fast(OPTIONAL) : Disable fail fast on build/tests/infra failures.--skip-test(OPTIONAL) : Skip all test stages, but still run build stages, package stages and sanity check stages. Note: Does NOT update GitHub check status.--stage-list "A10-PyTorch-1, xxx"(OPTIONAL) : Only run the specified test stages. Examples: "A10-PyTorch-1, xxx". Note: Does NOT update GitHub check status.--gpu-type "A30, H100_PCIe"(OPTIONAL) : Only run the test stages on the specified GPU types. Examples: "A30, H100_PCIe". Note: Does NOT update GitHub check status.--test-backend "pytorch, cpp"(OPTIONAL) : Skip test stages which don't match the specified backends. Only support [pytorch, cpp, tensorrt, triton]. Examples: "pytorch, cpp" (does not run test stages with tensorrt or triton backend). Note: Does NOT update GitHub pipeline status.--only-multi-gpu-test(OPTIONAL) : Only run the multi-GPU tests. Note: Does NOT update GitHub check status.--disable-multi-gpu-test(OPTIONAL) : Disable the multi-GPU tests. Note: Does NOT update GitHub check status.--add-multi-gpu-test(OPTIONAL) : Force run the multi-GPU tests in addition to running L0 pre-merge pipeline.--post-merge(OPTIONAL) : Run the L0 post-merge pipeline instead of the ordinary L0 pre-merge pipeline.--extra-stage "H100_PCIe-TensorRT-Post-Merge-1, xxx"(OPTIONAL) : Run the ordinary L0 pre-merge pipeline and specified test stages. Examples: --extra-stage "H100_PCIe-TensorRT-Post-Merge-1, xxx".--detailed-log(OPTIONAL) : Enable flushing out all logs to the Jenkins console. This will significantly increase the log volume and may slow down the job.--debug(OPTIONAL) : Experimental feature. Enable access to the CI container for debugging purpose. Note: Specify exactly one stage in thestage-listparameter to access the appropriate container environment. Note: Does NOT update GitHub check status.For guidance on mapping tests to stage names, see
docs/source/reference/ci-overview.mdand the
scripts/test_to_stage_mapping.pyhelper.kill
killKill all running builds associated with pull request.
skip
skip --comment COMMENTSkip testing for latest commit on pull request.
--comment "Reason for skipping build/test"is required. IMPORTANT NOTE: This is dangerous since lack of user care and validation can cause top of tree to break.reuse-pipeline
reuse-pipelineReuse a previous pipeline to validate current commit. This action will also kill all currently running builds associated with the pull request. IMPORTANT NOTE: This is dangerous since lack of user care and validation can cause top of tree to break.