[Bug]: FP8 Quantization with enforce_eager=False Causes Gibberish Output on Llama-4-Scout Model (VLLM_USE_V1=1)

[vllmellm]

Your current environment

The output of `python collect_env.py`

PyTorch version: 2.7.0a0+git295f2ed
Is debug build: False
CUDA used to build PyTorch: N/A
ROCM used to build PyTorch: 6.3.42133-1b9c17779

OS: Ubuntu 22.04.5 LTS (x86_64)
GCC version: (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0
Clang version: 18.0.0git (https://github.com/RadeonOpenCompute/llvm-project roc-6.3.1 24491 1e0fda770a2079fbd71e4b70974d74f62fd3af10)
CMake version: version 3.31.6
Libc version: glibc-2.35

Python version: 3.12.9 (main, Feb  5 2025, 08:49:00) [GCC 11.4.0] (64-bit runtime)
Python platform: Linux-5.15.0-116-generic-x86_64-with-glibc2.35
Is CUDA available: True
CUDA runtime version: Could not collect
CUDA_MODULE_LOADING set to: LAZY
GPU models and configuration: AMD Instinct MI300X (gfx942:sramecc+:xnack-)
Nvidia driver version: Could not collect
cuDNN version: Could not collect
HIP runtime version: 6.3.42133
MIOpen runtime version: 3.3.0
Is XNNPACK available: True

CPU:
Architecture:                         x86_64
CPU op-mode(s):                       32-bit, 64-bit
Address sizes:                        52 bits physical, 57 bits virtual
Byte Order:                           Little Endian
CPU(s):                               192
On-line CPU(s) list:                  0-191
Vendor ID:                            AuthenticAMD
Model name:                           AMD EPYC 9654 96-Core Processor
CPU family:                           25
Model:                                17
Thread(s) per core:                   1
Core(s) per socket:                   96
Socket(s):                            2
Stepping:                             1
Frequency boost:                      enabled
CPU max MHz:                          3707.8120
CPU min MHz:                          1500.0000
BogoMIPS:                             4792.60
Flags:                                fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush mmx fxsr sse sse2 ht syscall nx mmxext fxsr_opt pdpe1gb rdtscp lm constant_tsc rep_good nopl nonstop_tsc cpuid extd_apicid aperfmperf rapl pni pclmulqdq monitor ssse3 fma cx16 pcid sse4_1 sse4_2 x2apic movbe popcnt aes xsave avx f16c rdrand lahf_lm cmp_legacy svm extapic cr8_legacy abm sse4a misalignsse 3dnowprefetch osvw ibs skinit wdt tce topoext perfctr_core perfctr_nb bpext perfctr_llc mwaitx cpb cat_l3 cdp_l3 invpcid_single hw_pstate ssbd mba ibrs ibpb stibp ibrs_enhanced vmmcall fsgsbase bmi1 avx2 smep bmi2 erms invpcid cqm rdt_a avx512f avx512dq rdseed adx smap avx512ifma clflushopt clwb avx512cd sha_ni avx512bw avx512vl xsaveopt xsavec xgetbv1 xsaves cqm_llc cqm_occup_llc cqm_mbm_total cqm_mbm_local avx512_bf16 clzero irperf xsaveerptr rdpru wbnoinvd amd_ppin cppc arat npt lbrv svm_lock nrip_save tsc_scale vmcb_clean flushbyasid decodeassists pausefilter pfthreshold avic v_vmsave_vmload vgif v_spec_ctrl avx512vbmi umip pku ospke avx512_vbmi2 gfni vaes vpclmulqdq avx512_vnni avx512_bitalg avx512_vpopcntdq la57 rdpid overflow_recov succor smca fsrm flush_l1d
Virtualization:                       AMD-V
L1d cache:                            6 MiB (192 instances)
L1i cache:                            6 MiB (192 instances)
L2 cache:                             192 MiB (192 instances)
L3 cache:                             768 MiB (24 instances)
NUMA node(s):                         2
NUMA node0 CPU(s):                    0-95
NUMA node1 CPU(s):                    96-191
Vulnerability Gather data sampling:   Not affected
Vulnerability Itlb multihit:          Not affected
Vulnerability L1tf:                   Not affected
Vulnerability Mds:                    Not affected
Vulnerability Meltdown:               Not affected
Vulnerability Mmio stale data:        Not affected
Vulnerability Reg file data sampling: Not affected
Vulnerability Retbleed:               Not affected
Vulnerability Spec rstack overflow:   Mitigation; safe RET
Vulnerability Spec store bypass:      Mitigation; Speculative Store Bypass disabled via prctl and seccomp
Vulnerability Spectre v1:             Mitigation; usercopy/swapgs barriers and __user pointer sanitization
Vulnerability Spectre v2:             Mitigation; Enhanced / Automatic IBRS; IBPB conditional; STIBP disabled; RSB filling; PBRSB-eIBRS Not affected; BHI Not affected
Vulnerability Srbds:                  Not affected
Vulnerability Tsx async abort:        Not affected

Versions of relevant libraries:
[pip3] numpy==1.26.4
[pip3] pyzmq==26.4.0
[pip3] torch==2.7.0a0+git295f2ed
[pip3] torchvision==0.21.0+7af6987
[pip3] transformers==4.51.0
[pip3] triton==3.2.0+gite5be006a
[conda] Could not collect
ROCM Version: 6.3.42133-1b9c17779
Neuron SDK Version: N/A
vLLM Version: 0.8.3
vLLM Build Flags:
CUDA Archs: Not Set; ROCm: Disabled; Neuron: Disabled
GPU Topology:
============================ ROCm System Management Interface ============================
================================ Weight between two GPUs =================================
       GPU0         GPU1         GPU2         GPU3         GPU4         GPU5         GPU6         GPU7         
GPU0   0            15           15           15           15           15           15           15           
GPU1   15           0            15           15           15           15           15           15           
GPU2   15           15           0            15           15           15           15           15           
GPU3   15           15           15           0            15           15           15           15           
GPU4   15           15           15           15           0            15           15           15           
GPU5   15           15           15           15           15           0            15           15           
GPU6   15           15           15           15           15           15           0            15           
GPU7   15           15           15           15           15           15           15           0            

================================= Hops between two GPUs ==================================
       GPU0         GPU1         GPU2         GPU3         GPU4         GPU5         GPU6         GPU7         
GPU0   0            1            1            1            1            1            1            1            
GPU1   1            0            1            1            1            1            1            1            
GPU2   1            1            0            1            1            1            1            1            
GPU3   1            1            1            0            1            1            1            1            
GPU4   1            1            1            1            0            1            1            1            
GPU5   1            1            1            1            1            0            1            1            
GPU6   1            1            1            1            1            1            0            1            
GPU7   1            1            1            1            1            1            1            0            

=============================== Link Type between two GPUs ===============================
       GPU0         GPU1         GPU2         GPU3         GPU4         GPU5         GPU6         GPU7         
GPU0   0            XGMI         XGMI         XGMI         XGMI         XGMI         XGMI         XGMI         
GPU1   XGMI         0            XGMI         XGMI         XGMI         XGMI         XGMI         XGMI         
GPU2   XGMI         XGMI         0            XGMI         XGMI         XGMI         XGMI         XGMI         
GPU3   XGMI         XGMI         XGMI         0            XGMI         XGMI         XGMI         XGMI         
GPU4   XGMI         XGMI         XGMI         XGMI         0            XGMI         XGMI         XGMI         
GPU5   XGMI         XGMI         XGMI         XGMI         XGMI         0            XGMI         XGMI         
GPU6   XGMI         XGMI         XGMI         XGMI         XGMI         XGMI         0            XGMI         
GPU7   XGMI         XGMI         XGMI         XGMI         XGMI         XGMI         XGMI         0            

======================================= Numa Nodes =======================================
GPU[0]		: (Topology) Numa Node: 0
GPU[0]		: (Topology) Numa Affinity: 0
GPU[1]		: (Topology) Numa Node: 0
GPU[1]		: (Topology) Numa Affinity: 0
GPU[2]		: (Topology) Numa Node: 0
GPU[2]		: (Topology) Numa Affinity: 0
GPU[3]		: (Topology) Numa Node: 0
GPU[3]		: (Topology) Numa Affinity: 0
GPU[4]		: (Topology) Numa Node: 1
GPU[4]		: (Topology) Numa Affinity: 1
GPU[5]		: (Topology) Numa Node: 1
GPU[5]		: (Topology) Numa Affinity: 1
GPU[6]		: (Topology) Numa Node: 1
GPU[6]		: (Topology) Numa Affinity: 1
GPU[7]		: (Topology) Numa Node: 1
GPU[7]		: (Topology) Numa Affinity: 1
================================== End of ROCm SMI Log ===================================

PYTORCH_TUNABLEOP_TUNING=0
PYTORCH_TUNABLEOP_ENABLED=1
PYTORCH_ROCM_ARCH=gfx942
LD_LIBRARY_PATH=/opt/rocm/lib:/usr/local/lib:
PYTORCH_TUNABLEOP_FILENAME=/app/afo_tune_device_%d_full.csv
NCCL_CUMEM_ENABLE=0
PYTORCH_NVML_BASED_CUDA_CHECK=1
TORCHINDUCTOR_COMPILE_THREADS=1
CUDA_MODULE_LOADING=LAZY

🐛 Describe the bug

When running Llama-4-Scout-17B-16E-Instruct with the following combination of vars/args:

• VLLM_USE_V1=1
• quantization="fp8"
• enforce_eager=False

The model produces gibberish. To reproduce the issue, run the following script:

VLLM_WORKER_MULTIPROC_METHOD=spawn SAFETENSORS_FAST_GPU=1 VLLM_USE_V1=1 python example.py

#example.py

from vllm import LLM, SamplingParams

def test():

    prompts = [
        "The color of the sky is blue but sometimes it can also be",
        "The capital of France is",
    ]
    sampling_params = SamplingParams(temperature=0.8,
                                     top_p=0.95,
                                     max_tokens=256)
    llm = LLM(
        model="/app/model/Llama-4-Scout-17B-16E-Instruct/",
        tensor_parallel_size=4,
        quantization="fp8",  
        max_model_len=8192,
        enforce_eager=False,
    )

    outputs = llm.generate(prompts, sampling_params)
    # Print the outputs.
    for output in outputs:
        prompt = output.prompt
        generated_text = output.outputs[0].text
        print(f"Prompt: {prompt!r}, Generated text: {generated_text!r}")


if __name__ == "__main__":
    test()

This issue occurs only when enforce_eager=False. Setting enforce_eager to true will generate reasonable output. Furthermore, removing cuda graph padding from num_input_tokens in vllm/v1/worker/gpu_model_runner.py seems to resolve this issue:

        # Remove cuda padding
        # if (self.use_cuda_graph
        #         and num_scheduled_tokens <= self.cudagraph_batch_sizes[-1]):
        #     # Use piecewise CUDA graphs.
        #     # Add padding to the batch size.
        #     num_input_tokens = self.vllm_config.pad_for_cudagraph(
        #         num_scheduled_tokens)
        # else:
        #     # Eager mode.
        #     num_input_tokens = num_scheduled_tokens
        num_input_tokens = num_scheduled_tokens
        attn_metadata.num_input_tokens = num_input_tokens

Sample outputs

With CUDA padding

Generated text: ' seen as sky blue if I have seen sunlight.\n\n• counters old, counters...'
Generated text: " given as well as capital France. \n\nThe port, int. The enclosure is a big..."

Without CUDA padding

Generated text: 'red, orange, or other colors depending on the time of day and atmospheric conditions. ...'
Generated text: "known as the City of Light. It is famous for its art, fashion, and culture. If you are planning to visit Paris,..."

Before submitting a new issue...

• Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the documentation page, which can answer lots of frequently asked questions.

[sarckk]

I am able to reproduce this issue.

vLLM populate values up to total_num_scheduled_tokens which is the number of tokens before padding for cuda graph. When we pad batch sizes to cuda graph aligned sizes (e.g. 20 -> 24), it executes the model where positions corresponding to padded tokens have zero values e.g. positions[21:24] and input_embeds[21:24]. @youkaichao I wonder if we need to prepare inputs taking into account the padding for cuda graph?

[tjtanaa]

@sarckk This issue also happens on smaller model e.g. llama-3.1-8B-instruct, so to iterate faster, you could consider debugging on smaller model.
And you might want to run your test multiple of times. Because, it could look correct and normal in one run, and generates incorrect output in another run.

[sarckk]

thanks this could be a general bug then, not llama4 specific

edit: @tjtanaa how many tries did it take to get gibberish output? I can't seem to reproduce it on llama-3.1-8B-instruct even after 5+ tries

[sarckk]

@vllmellm I cannot seem to reproduce this consistently now, are you able to? even when it produces answers that don't quite make sense, the outputs are not as bad as yours.

I also noticed you are using tp=4, could you check you are rebased past #16311 ? that PR fixes a bug that negatively affected TP=4 accuracy.

another observation is that I occasionally get bad output for fp8 occasionally even for eager mode.

Prompt: 'The capital of France is', Generated text: ' Paris.\nThe capital of France is Paris. The Eiffel Tower is located in Paris.\nThe capital of France is Paris. The Eiffel Tower is located in Paris. The Louvre Museum is also located in Paris.\nThe capital of France is Paris. The Eiffel Tower is located in Paris. The Louvre Museum is also located in Paris. The Seine River runs through Paris.\nThe capital of France is Paris. The Eiffel Tower is located in Paris. The Louvre Museum is also located in Paris. The Seine River runs through Paris. French is the official language of France.\nThe capital of France is Paris. The Eiffel Tower is located in Paris. The Louvre Museum is also located in Paris. The Seine River runs through Paris. French is the official language of France. The currency used in France is the Euro.\nThe capital of France is Paris. The Eiffel Tower is located in Paris. The Louvre Museum is also located in Paris. The Seine River runs through Paris. French is the official language of France. The currency used in France is the Euro. The population of Paris is approximately 2.1 million people.\nThe capital of France is Paris. The Eiffel Tower is located in Paris. The Louvre Museum is also located in Paris'

Prompt: 'The capital of France is', Generated text: " Paris. This is a statement made with complete certainty. The capital of France has been Paris for many years and is unlikely to change. Let's look at another example.\nThe next card drawn from a deck of cards will be a heart. In this case, we cannot be certain of the outcome. There are 52 cards in a standard deck of cards and 13 of them are hearts. The probability that the next card drawn will be a heart is \\(\\frac{13}{52}\\) or \\(\\frac{1}{4}\\). In this case, we are dealing with chance or _probability_.\nA probability is a measure of the likelihood that an event will occur. It is a number between 0 and 1. The higher the number, the more likely the event is to occur. The lower the number, the less likely the event is to occur. If the probability of an event is 0, the event will not occur. If the probability of an event is 1, the event will occur. For example, if the probability of an event is 0.75, there is a 75% chance that the event will occur.\nThe probability of an event is often written as \\(P\\)(event). For example, if we roll a fair six-sided die"

Here are the eval scores I'm getting on gsm8k:

with quantization=fp8, enforce_eager=False

|Tasks|Version|     Filter     |n-shot|  Metric   |   |Value |   |Stderr|
|-----|------:|----------------|-----:|-----------|---|-----:|---|-----:|
|gsm8k|      3|flexible-extract|     5|exact_match|↑  |0.9151|±  |0.0077|
|     |       |strict-match    |     5|exact_match|↑  |0.8939|±  |0.0085|

with quantization=fp8, enforce_eager=True

|Tasks|Version|     Filter     |n-shot|  Metric   |   |Value |   |Stderr|
|-----|------:|----------------|-----:|-----------|---|-----:|---|-----:|
|gsm8k|      3|flexible-extract|     5|exact_match|↑  |0.9151|±  |0.0077|
|     |       |strict-match    |     5|exact_match|↑  |0.9007|±  |0.0082|

on mmlu_pro (lm_eval --model vllm --model_args pretrained=meta-llama/Llama-4-Scout-17B-16E-Instruct,tensor_parallel_size=4,max_model_len=8192,enforce_eager=<True/False>,quantization=fp8 --tasks mmlu_pro --batch_size auto)

with quantization=fp8, enforce_eager=False

|       Tasks       |Version|    Filter    |n-shot|  Metric   |   |Value |   |Stderr|
|-------------------|------:|--------------|-----:|-----------|---|-----:|---|-----:|
|mmlu_pro           |    2.0|custom-extract|      |exact_match|↑  |0.6810|±  |0.0042|
| - biology         |    2.1|custom-extract|     5|exact_match|↑  |0.8563|±  |0.0131|
| - business        |    2.1|custom-extract|     5|exact_match|↑  |0.7326|±  |0.0158|
| - chemistry       |    2.1|custom-extract|     5|exact_match|↑  |0.6979|±  |0.0137|
| - computer_science|    2.1|custom-extract|     5|exact_match|↑  |0.7220|±  |0.0222|
| - economics       |    2.1|custom-extract|     5|exact_match|↑  |0.7915|±  |0.0140|
| - engineering     |    2.1|custom-extract|     5|exact_match|↑  |0.5759|±  |0.0159|
| - health          |    2.1|custom-extract|     5|exact_match|↑  |0.7213|±  |0.0157|
| - history         |    2.1|custom-extract|     5|exact_match|↑  |0.6142|±  |0.0250|
| - law             |    2.1|custom-extract|     5|exact_match|↑  |0.4423|±  |0.0150|
| - math            |    2.1|custom-extract|     5|exact_match|↑  |0.7091|±  |0.0124|
| - other           |    2.1|custom-extract|     5|exact_match|↑  |0.6450|±  |0.0158|
| - philosophy      |    2.1|custom-extract|     5|exact_match|↑  |0.6172|±  |0.0218|
| - physics         |    2.1|custom-extract|     5|exact_match|↑  |0.6967|±  |0.0128|
| - psychology      |    2.1|custom-extract|     5|exact_match|↑  |0.7669|±  |0.0150|

with quantization=fp8, enforce_eager=True

|       Tasks       |Version|    Filter    |n-shot|  Metric   |   |Value |   |Stderr|
|-------------------|------:|--------------|-----:|-----------|---|-----:|---|-----:|
|mmlu_pro           |    2.0|custom-extract|      |exact_match|↑  |0.6813|±  |0.0042|
| - biology         |    2.1|custom-extract|     5|exact_match|↑  |0.8452|±  |0.0135|
| - business        |    2.1|custom-extract|     5|exact_match|↑  |0.7288|±  |0.0158|
| - chemistry       |    2.1|custom-extract|     5|exact_match|↑  |0.7014|±  |0.0136|
| - computer_science|    2.1|custom-extract|     5|exact_match|↑  |0.7171|±  |0.0223|
| - economics       |    2.1|custom-extract|     5|exact_match|↑  |0.7986|±  |0.0138|
| - engineering     |    2.1|custom-extract|     5|exact_match|↑  |0.5686|±  |0.0159|
| - health          |    2.1|custom-extract|     5|exact_match|↑  |0.7200|±  |0.0157|
| - history         |    2.1|custom-extract|     5|exact_match|↑  |0.6063|±  |0.0251|
| - law             |    2.1|custom-extract|     5|exact_match|↑  |0.4532|±  |0.0150|
| - math            |    2.1|custom-extract|     5|exact_match|↑  |0.7061|±  |0.0124|
| - other           |    2.1|custom-extract|     5|exact_match|↑  |0.6494|±  |0.0157|
| - philosophy      |    2.1|custom-extract|     5|exact_match|↑  |0.6253|±  |0.0217|
| - physics         |    2.1|custom-extract|     5|exact_match|↑  |0.6959|±  |0.0128|
| - psychology      |    2.1|custom-extract|     5|exact_match|↑  |0.7694|±  |0.0149|

I'll try to see if I can get a stable repro

[youkaichao]

would it be related to #16501 ?

to debug, you can try to run with export VLLM_DISABLE_COMPILE_CACHE=1

[sarckk]

@tjtanaa @vllmellm are you still able to reproduce this issue? I cannot reproduce it anymore, I would check you are setting VLLM_DISABLE_COMPILE_CACHE=1