[NVIDIA] Set EP_SIZE = 1 for B200 measurements

[kaixih]

This PR disable the ep for B200 measurements.

[Copilot]

Pull Request Overview

This PR disables endpoint parallelism (EP) for B200 measurements by setting EP_SIZE to 1 across all configurations. The change affects both FP4 and FP8 precision test configurations for the B200 platform using sglang.

• Changed EP values from matching TP values (4 or 8) to a uniform value of 1
• Applied consistently across all input/output sequence length combinations

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[cquil11]

does this lead to perf improvement?

[kaixih]

does this lead to perf improvement?

I think a previous commit enables the EP accidentally. This PR tries to recover it.

[functionstackx]

@kaixih wasn't EP always enabled?

before PR https://github.com/InferenceMAX/InferenceMAX/pull/204 , --enable-ep-moe & --enable-flashinfer-trtllm-moe was set

according to SGLang docs, since --enable-ep-moe flag has been removed, the equivalent is to set it as --ep-size "Please set --ep-size to the same value as --tp-size instead"

[Copilot]

Pull Request Overview

Copilot reviewed 1 out of 1 changed files in this pull request and generated no new comments.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[kaixih]

@functionstackx Right, I re-enabled the EP for the fp4 workloads. We noticed that the EP was off only for FP8 and was accidentally enabled from this InferenceMAX/InferenceMAX@d8fe8f7.

[functionstackx]

@functionstackx Right, I re-enabled the EP for the fp4 workloads. We noticed that the EP was off only for FP8 and was accidentally enabled from this d8fe8f7.

@kaixih ah gotcha, i see. it was only fp4 that previously had --enable-ep-moe and fp8 didn't have --enable-ep-moe

I see now

[functionstackx]

lgtm