[BugFix] MLA + V1, illegal memory access and accuracy issues

[LucasWilkinson]

Fix the decode / prefill split in V1 with MLA not being handled correctly leading to inaccurate results when max_num_batched_tokens is low and illegal memory accesses. Also refactored the code to precompute the tensor slicing for prefill and decode metadata.

Main:

** max_num_batched_tokens=1024

VLLM_USE_V1=1 lm_eval --model vllm --model_args pretrained=deepseek-ai/DeepSeek-V2-Lite-Chat,tensor_parallel_size=2,dtype=auto,gpu_memory_utilization=0.9,trust_remote_code=True,max_model_len=16384,enforce_eager=True,max_num_batched_tokens=1024 --task gsm8k --num_fewshot=5 --limit 10

|Tasks|Version|     Filter     |n-shot|  Metric   |   |Value|   |Stderr|
|-----|------:|----------------|-----:|-----------|---|----:|---|-----:|
|gsm8k|      3|flexible-extract|     5|exact_match|↑  |  0.5|±  |0.1667|
|     |       |strict-match    |     5|exact_match|↑  |  0.5|±  |0.1667|


** max_num_batched_tokens=Default

VLLM_USE_V1=1 lm_eval --model vllm --model_args pretrained=deepseek-ai/DeepSeek-V2-Lite-Chat,tensor_parallel_size=2,dtype=auto,gpu_memory_utilization=0.9,
trust_remote_code=True,max_model_len=16384,enforce_eager=True --task gsm8k --num_fewshot=5 --limit 10

|Tasks|Version|     Filter     |n-shot|  Metric   |   |Value|   |Stderr|
|-----|------:|----------------|-----:|-----------|---|----:|---|-----:|
|gsm8k|      3|flexible-extract|     5|exact_match|↑  |  0.8|±  |0.1333|
|     |       |strict-match    |     5|exact_match|↑  |  0.8|±  |0.1333|

This PR:

** max_num_batched_tokens=1024

VLLM_USE_V1=1 lm_eval --model vllm --model_args pretrained=deepseek-ai/DeepSeek-V2-Lite-Chat,tensor_parallel_size=2,dtype=auto,gpu_memory_utilization=0.9,trust_remote_code=True,max_model_len=16384,enforce_eager=True,max_num_batched_tokens=1024 --task gsm8k --num_fewshot=5 --limit 10

|Tasks|Version|     Filter     |n-shot|  Metric   |   |Value|   |Stderr|
|-----|------:|----------------|-----:|-----------|---|----:|---|-----:|
|gsm8k|      3|flexible-extract|     5|exact_match|↑  |  0.8|±  |0.1333|
|     |       |strict-match    |     5|exact_match|↑  |  0.8|±  |0.1333|

VLLM_USE_V1=1 VLLM_ATTENTION_BACKEND=FLASHMLA  lm_eval --model vllm --model_args pretrained=deepseek-ai/DeepSeek-V2-Lite-Chat,tensor_parallel_size=2,dtype=auto,gpu_memory_utilization=0.9,trust_remote_code=True,max_model_len=16384,enforce_eager=True,max_num_batched_tokens=1024 --task gsm8k --num_fewshot=5 --limit 10

|Tasks|Version|     Filter     |n-shot|  Metric   |   |Value|   |Stderr|
|-----|------:|----------------|-----:|-----------|---|----:|---|-----:|
|gsm8k|      3|flexible-extract|     5|exact_match|↑  |  0.8|±  |0.1333|
|     |       |strict-match    |     5|exact_match|↑  |  0.8|±  |0.1333|

[github-actions]

👋 Hi! Thank you for contributing to the vLLM project.

💬 Join our developer Slack at https://slack.vllm.ai to discuss your PR in #pr-reviews, coordinate on features in #feat- channels, or join special interest groups in #sig- channels.

Just a reminder: PRs would not trigger full CI run by default. Instead, it would only run fastcheck CI which starts running only a small and essential subset of CI tests to quickly catch errors. You can run other CI tests on top of those by going to your fastcheck build on Buildkite UI (linked in the PR checks section) and unblock them. If you do not have permission to unblock, ping simon-mo or khluu to add you in our Buildkite org.

Once the PR is approved and ready to go, your PR reviewer(s) can run CI to test the changes comprehensively before merging.

To run CI, PR reviewers can either: Add ready label to the PR or enable auto-merge.

🚀

[mgoin]

Unfortunately this doesn't seem to fix full R1 evals. I ran GSM8k on 8xH200

V0 works fine:

(vllm) ➜  vllm git:(lwilkinson/fix-mla-v1-ima) ✗ lm_eval --model vllm --model_args pretrained=/home/vllm-dev/DeepSeek-R1,tensor_parallel_size=8,max_model_len=4096 --trust_remote_code --tasks gsm8k --num_fewshot 5 --batch_size auto     
...
vllm (pretrained=/home/vllm-dev/DeepSeek-R1,tensor_parallel_size=8,max_model_len=4096,trust_remote_code=True), gen_kwargs: (None), limit: None, num_fewshot: 5, batch_size: auto
|Tasks|Version|     Filter     |n-shot|  Metric   |   |Value |   |Stderr|
|-----|------:|----------------|-----:|-----------|---|-----:|---|-----:|
|gsm8k|      3|flexible-extract|     5|exact_match|↑  |0.9507|±  | 0.006|
|     |       |strict-match    |     5|exact_match|↑  |0.9492|±  | 0.006|

V1 fails to get a good score (I get the same for VLLM_ATTENTION_BACKEND=FLASHMLA as well):

(vllm) ➜  vllm git:(lwilkinson/fix-mla-v1-ima) ✗ VLLM_USE_V1=1 lm_eval --model vllm --model_args pretrained=/home/vllm-dev/DeepSeek-R1,tensor_parallel_size=8,max_model_len=4096 --trust_remote_code --tasks gsm8k --num_fewshot 5 --batch_size auto
...
vllm (pretrained=/home/vllm-dev/DeepSeek-R1,tensor_parallel_size=8,max_model_len=4096,trust_remote_code=True), gen_kwargs: (None), limit: None, num_fewshot: 5, batch_size: auto
|Tasks|Version|     Filter     |n-shot|  Metric   |   |Value |   |Stderr|
|-----|------:|----------------|-----:|-----------|---|-----:|---|-----:|
|gsm8k|      3|flexible-extract|     5|exact_match|↑  |0.2358|±  |0.0117|
|     |       |strict-match    |     5|exact_match|↑  |0.2358|±  |0.0117|

[LucasWilkinson]

Fixed 👍

VLLM_USE_V1=1 lm_eval --model vllm --model_args pretrained=/home/vllm-dev/DeepSeek-R1,tensor_parallel_size=8,max_model_len=4096 --trust_remote_code --tasks gsm8k --num_fewshot 5 --batch_size auto

|Tasks|Version|     Filter     |n-shot|  Metric   |   |Value |   |Stderr|
|-----|------:|----------------|-----:|-----------|---|-----:|---|-----:|
|gsm8k|      3|flexible-extract|     5|exact_match|↑  |0.9522|±  |0.0059|
|     |       |strict-match    |     5|exact_match|↑  |0.9515|±  |0.0059|

VLLM_USE_V1=1 VLLM_ATTENTION_BACKEND=FLASHMLA lm_eval --model vllm --model_args pretrained=/home/vllm-dev/DeepSeek-R1,tensor_parallel_size=8,max
_model_len=4096 --trust_remote_code --tasks gsm8k --num_fewshot 5 --batch_size auto

|Tasks|Version|     Filter     |n-shot|  Metric   |   |Value |   |Stderr|
|-----|------:|----------------|-----:|-----------|---|-----:|---|-----:|
|gsm8k|      3|flexible-extract|     5|exact_match|↑  |0.9507|±  | 0.006|
|     |       |strict-match    |     5|exact_match|↑  |0.9500|±  | 0.006|

[mgoin]

Amazing work solving this so fast! 🔥🔥🔥

[pathorn]

With this bugfix, V1 now seems stable on DeepSeek R1 tp=8 H200 both with triton and FlashMLA attn backend.
Nice work!