llm: set ROCm Tensile lib path by HEETMEHTA18 · Pull Request #15527 · ollama/ollama

HEETMEHTA18 · 2026-04-12T19:32:55Z

Resolve rocBLAS Tensile path from bundled ROCm libs and pass it to the runner when unset. Add tests for path detection and env injection. rocblaslt error: Cannot read "TensileLibrary_lazy_gfx1200.dat": No such file or directory #15452

Summary

Fix ROCm Tensile library path resolution for runner startup on gfx1200-class systems.

On some ROCm setups, rocBLASLt first attempts a relative lookup for TensileLibrary_lazy_gfx1200.dat, logs a "No such file
or directory" warning, and then falls back to an absolute path. Inference may still work, but startup is noisy and path
resolution is fragile.

This patch makes lookup deterministic by resolving the Tensile library directory from gpuLibs and setting env vars at runner
startup when they are not already provided.

Changes

In llm/server.go:
- Add findRocmTensileLibraryPath(gpuLibs []string) string
- Add ensureRocmTensileEnv(gpuLibs []string, extraEnvs map[string]string) map[string]string
- Call ensureRocmTensileEnv(...) in StartRunner(...) before environment assembly
Inject (only when unset):
- ROCBLAS_TENSILE_LIBPATH
- ROCBLASLT_TENSILE_LIBPATH
- HIPBLASLT_TENSILE_LIBPATH
Supported candidate layouts:
- <lib>/rocblas/library
- <lib>/rocm/rocblas/library
Do not override user/process-provided env values.

Tests

Added unit tests in llm/server_test.go for:

Tensile path detection from both directory layouts
Injection behavior when env vars are unset
Non-overwrite behavior when env vars are explicitly set

Why this change

This addresses path-resolution fragility behind issue #15452 and removes noisy false-negative file lookup behavior in runner
startup.

Compatibility

No behavior change when env vars are already set by user/system.
No injection when a valid Tensile directory is not found.
Complements packaging-side fixes (e.g. hipBLASLt kernel packaging), but does not replace them.

Closes #15452

Resolve rocBLAS Tensile path from bundled ROCm libs and pass it to the runner when unset. Add tests for path detection and env injection.

Copilot

Pull request overview

This PR makes ROCm runner startup more deterministic by locating the bundled rocBLAS Tensile library directory from gpuLibs and injecting Tensile-related env vars into the runner environment when they aren’t already provided, reducing noisy rocblaslt “Cannot read TensileLibrary…” warnings on gfx1200-class systems.

Changes:

Add ROCm Tensile library path detection from common bundled ROCm layouts.
Inject ROCBLAS*_TENSILE_LIBPATH / HIPBLASLT_TENSILE_LIBPATH into the runner env when not already set.
Add unit tests covering path detection and env injection / non-overwrite behavior.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 2 comments.

File	Description
llm/server.go	Adds Tensile path detection and env injection during `StartRunner`.
llm/server_test.go	Adds unit tests validating Tensile path detection and env injection behavior.

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

Copilot · 2026-04-12T19:36:38Z

+	for _, key := range []string{"ROCBLAS_TENSILE_LIBPATH", "ROCBLASLT_TENSILE_LIBPATH", "HIPBLASLT_TENSILE_LIBPATH"} {
+		if _, ok := os.LookupEnv(key); ok {
+			return extraEnvs
+		}
+		if extraEnvs != nil {
+			if _, ok := extraEnvs[key]; ok {
+				return extraEnvs
+			}
+		}
+	}
+
+	if extraEnvs == nil {
+		extraEnvs = map[string]string{}
+	}
+
+	extraEnvs["ROCBLAS_TENSILE_LIBPATH"] = tensilePath
+	extraEnvs["ROCBLASLT_TENSILE_LIBPATH"] = tensilePath
+	extraEnvs["HIPBLASLT_TENSILE_LIBPATH"] = tensilePath
+


ensureRocmTensileEnv currently returns early if any of the Tensile env vars are set in the process env or extraEnvs, which means it will not inject the remaining vars when only one is provided. If (for example) ROCBLAS_TENSILE_LIBPATH is set but ROCBLASLT_TENSILE_LIBPATH is not, the rocBLASLt startup warning can still occur. Consider setting each of the three variables independently when it is not already present (leaving any explicitly provided values untouched).

Suggested change

for _, key := range []string{"ROCBLAS_TENSILE_LIBPATH", "ROCBLASLT_TENSILE_LIBPATH", "HIPBLASLT_TENSILE_LIBPATH"} {

if _, ok := os.LookupEnv(key); ok {

return extraEnvs

}

if extraEnvs != nil {

if _, ok := extraEnvs[key]; ok {

return extraEnvs

}

}

}

if extraEnvs == nil {

extraEnvs = map[string]string{}

}

extraEnvs["ROCBLAS_TENSILE_LIBPATH"] = tensilePath

extraEnvs["ROCBLASLT_TENSILE_LIBPATH"] = tensilePath

extraEnvs["HIPBLASLT_TENSILE_LIBPATH"] = tensilePath

if extraEnvs == nil {

extraEnvs = map[string]string{}

}

for _, key := range []string{"ROCBLAS_TENSILE_LIBPATH", "ROCBLASLT_TENSILE_LIBPATH", "HIPBLASLT_TENSILE_LIBPATH"} {

if _, ok := os.LookupEnv(key); ok {

continue

}

if _, ok := extraEnvs[key]; ok {

continue

}

extraEnvs[key] = tensilePath

}

Copilot · 2026-04-12T19:36:38Z

+	t.Run("do not override explicit extra env", func(t *testing.T) {
+		in := map[string]string{"ROCBLASLT_TENSILE_LIBPATH": "/custom/path"}
+		envs := ensureRocmTensileEnv([]string{root}, in)
+		if envs["ROCBLASLT_TENSILE_LIBPATH"] != "/custom/path" {
+			t.Fatalf("ROCBLASLT_TENSILE_LIBPATH = %q, want /custom/path", envs["ROCBLASLT_TENSILE_LIBPATH"])
+		}
+		if _, ok := envs["ROCBLAS_TENSILE_LIBPATH"]; ok {
+			t.Fatalf("ROCBLAS_TENSILE_LIBPATH should not be injected when explicit env is set")
+		}
+	})


The test case "do not override explicit extra env" asserts that no other Tensile env vars are injected when one (ROCBLASLT_TENSILE_LIBPATH) is explicitly set. If the intended behavior is to inject only the missing vars (without overriding the explicitly-set one), this test should be updated to expect injection of the other keys rather than blocking all injection.

Copilot

Pull request overview

Copilot reviewed 2 out of 2 changed files in this pull request and generated 2 comments.

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

Copilot · 2026-04-12T19:44:51Z

+		if _, ok := os.LookupEnv(key); ok {
+			continue
+		}
+		if extraEnvs != nil {
+			if _, ok := extraEnvs[key]; ok {


ensureRocmTensileEnv treats an environment variable as “set” purely based on os.LookupEnv. If any of these vars exist but are an empty string (common in some service/unit configs), the helper will skip injection and the rocBLASLt relative lookup noise can still occur. Consider treating empty values as unset (e.g., check val != "" for both process env and extraEnvs).

Suggested change

if _, ok := os.LookupEnv(key); ok {

continue

}

if extraEnvs != nil {

if _, ok := extraEnvs[key]; ok {

if val, ok := os.LookupEnv(key); ok && val != "" {

continue

}

if extraEnvs != nil {

if val, ok := extraEnvs[key]; ok && val != "" {

Copilot · 2026-04-12T19:44:51Z

+	t.Run("inject when unset", func(t *testing.T) {
+		envs := ensureRocmTensileEnv([]string{root}, nil)
+		if envs["ROCBLAS_TENSILE_LIBPATH"] != tensile {
+			t.Fatalf("ROCBLAS_TENSILE_LIBPATH = %q, want %q", envs["ROCBLAS_TENSILE_LIBPATH"], tensile)
+		}
+		if envs["ROCBLASLT_TENSILE_LIBPATH"] != tensile {
+			t.Fatalf("ROCBLASLT_TENSILE_LIBPATH = %q, want %q", envs["ROCBLASLT_TENSILE_LIBPATH"], tensile)
+		}
+		if envs["HIPBLASLT_TENSILE_LIBPATH"] != tensile {
+			t.Fatalf("HIPBLASLT_TENSILE_LIBPATH = %q, want %q", envs["HIPBLASLT_TENSILE_LIBPATH"], tensile)
+		}
+	})


These subtests assume ROCBLAS*_TENSILE_LIBPATH/HIPBLASLT_TENSILE_LIBPATH are unset in the parent process. On ROCm CI runners (or developer machines) they may already be set, causing the “inject when unset” assertions to fail. Make the test hermetic by explicitly clearing those env vars for the subtest (e.g., unset with cleanup, or set to empty if the helper treats empty-as-unset).

slojosic-amd · 2026-04-14T14:26:07Z

@HEETMEHTA18 Thanks for working on this — the underlying issue (#15452) is real and worth addressing.

One thing worth noting: ROCBLASLT_TENSILE_LIBPATH does not appear to correspond to any actual ROCm library. The ROCm stack ships librocblas.so and libhipblaslt.so, but there is no librocblaslt.so. The rocblaslt name visible in the hipBLASLt source tree (e.g. https://github.com/ROCm/rocm-libraries/tree/develop/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/Tensile/Logic/asm_full) is an internal implementation namespace, not a separate shared library and that's the reason why rocblaslt word can appear in error messages like these in #15452. As a result, setting ROCBLASLT_TENSILE_LIBPATH has no effect since nothing reads it.

However, the root cause of the missing TensileLibrary_lazy_gfx1200.dat error is that the hipblaslt/library/ folder was not being copied into the ollama install directory during the HIP CMake component install. A packaging-side fix that adds install(DIRECTORY ... hipblaslt ...) #14979 alongside the existing rocblas install rule would address this at the source rather than working around it at runtime so I think your PR would only add redundant env vars.

HEETMEHTA18 · 2026-04-14T14:59:33Z

@slojosic-amd Okaay ! I understood your reply.
But I agree packaging hipblaslt/library is the primary fix for Ollama’s bundled ROCm install. My intent with the
-llm change was to provide a runtime fallback for cases where ROCm assets are present but path discovery is
still brittle.

That said, if ROCBLASLT_TENSILE_LIBPATH is not actually consumed and the failing lookup is specifically due
to missing hipblaslt/library, then the current PR is too broad / aimed at the wrong place. I can rework this
into a packaging-only fix, or keep a much narrower runtime fallback only if we can point to a variable that
is actually honored in this path.

I could re work on this pr and revert the previous made change and add the right changes into this pr.

HEETMEHTA18 · 2026-04-16T08:23:47Z

@slojosic-amd i have added the changes in the CMakeList.txt with adding the required library to install to solve the issue and reverted all the old env changes which where temporary changes which i thought it would work but now this should be working now as from the ref of #14979.

llm: set ROCm Tensile lib path

1a669a2

Resolve rocBLAS Tensile path from bundled ROCm libs and pass it to the runner when unset. Add tests for path detection and env injection.

Copilot AI review requested due to automatic review settings April 12, 2026 19:32

Copilot started reviewing on behalf of HEETMEHTA18 April 12, 2026 19:33

View session

Copilot AI reviewed Apr 12, 2026

View reviewed changes

llm: fill missing ROCm Tensile env vars

3c09e42

HEETMEHTA18 requested review from Copilot April 12, 2026 19:42

Copilot started reviewing on behalf of HEETMEHTA18 April 12, 2026 19:43

View session

Copilot AI reviewed Apr 12, 2026

View reviewed changes

llm: treat empty Tensile env values as unset

c7702ed

rocm: package hipblaslt library

d7ba32e

Merge branch 'main' into fix/rocm-tensile-libpath-15452-clean

acd4394

Navigation Menu

Repository navigation

llm: set ROCm Tensile lib path - #15527

llm: set ROCm Tensile lib path#15527
HEETMEHTA18 wants to merge 5 commits into
ollama:mainfrom
HEETMEHTA18:fix/rocm-tensile-libpath-15452-clean

HEETMEHTA18 commented Apr 12, 2026

Uh oh!

Copilot AI left a comment

Uh oh!

Copilot AI Apr 12, 2026

Uh oh!

Copilot AI Apr 12, 2026

Uh oh!

Copilot AI left a comment

Uh oh!

Copilot AI Apr 12, 2026

Uh oh!

Copilot AI Apr 12, 2026

Uh oh!

slojosic-amd commented Apr 14, 2026 •

edited

Loading

Uh oh!

HEETMEHTA18 commented Apr 14, 2026

Uh oh!

HEETMEHTA18 commented Apr 16, 2026

Uh oh!

This branch has not been deployed

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

Conversation

HEETMEHTA18 commented Apr 12, 2026

Summary

Changes

Tests

Why this change

Compatibility

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Reviewed changes

Uh oh!

Copilot AI Apr 12, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Apr 12, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Uh oh!

Copilot AI Apr 12, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Apr 12, 2026

Choose a reason for hiding this comment

Uh oh!

slojosic-amd commented Apr 14, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

HEETMEHTA18 commented Apr 14, 2026

Uh oh!

HEETMEHTA18 commented Apr 16, 2026

Uh oh!

This branch has not been deployed

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

slojosic-amd commented Apr 14, 2026 •

edited

Loading